With the launch date of the RX 7000 cards now being so close, over the last few days, we have had several benchmark leaks. The first such leak showed the RX 7900 XTX severely underperforming, though, it was likely a relief for most enthusiasts and AMD fans to see a revised score in the same benchmark showing closer to a 100% improvement. This was more in line with what we estimated the performance of the 7900 XTX and XT to be using AMD-provided data.

Following the 7900 XTX revised leak, we also had the first leak of the smaller 7900 XT. While the performance was once again decent, it was quite a bit slower, by 16%, despite being just 10% cheaper. Overall this result may be somewhat of a disappointment for many owing to the high price of the XT GPU.

Following those Geekbench results, today we have another benchmark in the form of Blender. The Blender open database right now lists a single entry each for the 7900 XTX and 7900 XT. The two cards are now in reviewers' hands and it's possible one of them was testing the GPU on Blender in order to gauge what these cards can do.

Before we dive into the latest leak, bear in mind that these are still somewhat early drivers, even if they are the ones provided to the press. Hence it is possible that we could see a better showing with future driver updates. Also, this is one sample that is likely to be skewing the benchmark results and more test samples could certainly improve it or even make it worse. But we will still report on it and discuss it as we do with anything tech that is or can be, of interest to our readers. You are of course free to share your opinions below.

While AMD has not been great in Blender, the company is making decent progress gradually with Heterogeneous Interface for Portability (HIP) support improving with almost every new Blender release. However, it may not be enough to close the gap with Nvidia's OptiX or CUDA any time soon. The RX 7900 XTX and XT have respectively scored 3783.31 and 3461.95 points. Hence these are coming in at the level of last-gen RTX 3070 Ti or maybe a 3080 at best, at least according to these results.

The disappointing aspect though is the performance improvement of the two new AMD RDNA 3 GPUs compared to their predecessors. The RX 7900 XTX is only 34.8% and 41.2% faster than the 6950 XT and the 6900 XT respectively in the latest Blender 3.4. Meanwhile, the 7900 XT is 23.3% and 29.2% faster than the same two cards.

In the same test, the RX 6700 XT (RDNA 2) is around 60% faster than the 5700 XT (RDNA 1) card, and this is despite the 6700 XT having the same number of compute units (CUs) as the 5700 XT, unlike the 7900 XTX and XT that have more CUs than the 6900/6950 XT. Meanwhile, Nvidia RTX 4090 and 4080 show big gains as well compared to their previous-gen counterparts.

But as mentioned above, these are still not released and we could be looking at vastly improved performance figures in reviews in a few days. Besides, if you are a gamer only, this shouldn't bother you much.

Source: Blender via: Olrak (Twitter)

AMD RX 7900 XTX / XT new benchmark scores leak and they may not allure you

TSMC has seen impressive growth in its November sales to reach $7.27 billion in revenue. This puts the company on track to meet its fourth-quarter estimate of between $19.9 billion and $20.7 billion.

Dale Gai, a semiconductor analyst at Counterpoint Research, told CNBC:

“TSMC’s Oct/Nov revenues are on track comparing to what the management guided 2 months ago, despite significant business slowdowns in many other semi names"

The Taiwanese chipmaker's strong performance in the quarter is perhaps proportional to its supply of the A16 SoC used in the iPhone 14 Pro and 14 Pro Max. Apple, TSMC's biggest customer, announced that it would purchase chips from the supplier's US facilities, reducing its reliance on Asian manufacturers.

At an event attended by President Joe Biden and Apple CEO Tim Cook earlier this week, TSMC announced that it would expand its building and manufacturing efforts in the US. The expansion will involve a $40 billion investment in two fabs in Arizona, an increase from the previously planned $12 billion. The new facilities will also be able to begin manufacturing using 4nm process, which is an upgrade from the previously planned 5nm wafers.

The project is expected to create tens of thousands of jobs and have a total manufacturing output of 600,000 wafers per year.

Although TSMC's November revenue has been boosted by its partnership with Apple, analysts are concerned about potential weaker orders in the coming year. Despite this, TSMC reportedly plans to increase its 3nm wafer prices by 25% over 5nm to $20,000, which is expected to boost its bottom line. The 2023 iPhone 15's A17 Bionic chip will be manufactured using the 3nm process, and Apple has also agreed, albeit reluctantly, to the price hike.

Source: CNBC

Thanks to Apple, TSMC's November revenue soars by more than 50% yoy

The US should move expeditiously to reinforce its relationships with liberal-democratic allies

One significant point that Chris Miller (and, incidentally, many policymakers) seems to miss in his book Chip War is that discussions of semiconductors and national security are emphatically not solely, or even primarily, about the fabrication plants where chips are built.

One response to the China challenge is to persuade Taiwan Semiconductor Manufacturing Company (TSMC) or Samsung to build a modern fab in the US, and operate it as a foundry. But this would only ensure access to a specific process node (for example 5-nanometer or 7nm) that happens to be leading-edge today.

The US needs to ensure secure, asymmetric access to the world’s best chips across multiple generations. Building the world’s best chips is not a matter of a single factory or a single company: It’s a product of maintaining an entire semiconductor ecosystem, at home and within the borders of allied and partner regimes. 

There can be no chokepoints – fabrication and metrology tools, expert knowledge, chemicals, packaging, testing, design tools, or anything else – located solely in adversarial regimes. Maintaining semiconductor supremacy is about maintaining US and allied dominance across the entire ecosystem.

The liberal-democratic/Western semiconductor ecosystem still has enormous geographic stickiness and, in Halford Mackinder’s parlance, “social momentum.” Moving factories is enormously disruptive and expensive; engineers and executives don’t want to move their households and families to distant and culturally incompatible places. 

A critical question is how to leverage what remains of this social momentum in the US to enhance US dominance despite the fact that cutting-edge changes occur rapidly, across all areas of the semiconductor ecosystem. 

The US retains enormous momentum in the electronic design automation tools, the design of chips, and the development of fabrication and metrology tools. Taiwan has accumulated enormous momentum for the fabrication and packaging sides of the ecosystem, especially for cutting-edge digital chips, and is moving into design. 

China has very little momentum in most of these areas, but has made huge progress and constructed considerable momentum in trailing-edge and more-than-Moore chip fabrication technologies, as well as in integrating packaged chips on to boards and into finished products. 

The Communist Party of China (CPC) is spending money at an enormous pace to try to jumpstart ecosystems in all of these other areas.

Today’s most advanced chips from TSMC are in the 3nm node; if the US government pays TSMC to build a 3nm plant in the US today, the United States will have ensured access to 3nm chips. It will not, however, have secured access to the intellectual and organizational machinery that is constantly innovating and building newer and better fabrication processes and plants.

Taiwan’s social structure and its education system undergirds TSMC and that constant innovation, rolling out a new process every couple of years, is the core of TSMC’s long-term value.

The tools to implement these fabrication processes change too, and at the same pace. Every generation of new chips needs new fabrication equipment, new chemicals, new software, new metrology equipment, new test equipment, new higher-density printed circuit boards, new package designs, new design IP, etc. And every one of these pieces is a potential chokepoint for an adversarial regime, should it manage to wrest control of it. 

Consider the recent behavior of China with regard to the supply of rare-earth elements, a critical component in the development of many products relying on advanced technology: Upon establishing a monopoly on their supply through legal (and lethal) environmental externalities and government subsidies, the Chinese immediately made use of that leverage for political ends.

Some niches of the semiconductor ecosystem are actually natural monopolies. ASML has emerged as a monopoly because developing and building the EUV (extreme ultraviolet) steppers that are required for advanced node manufacturing is so ruinously expensive that ASML’s big customers were forced to directly subsidize their operations.

Demand for these machines, while large in dollar terms, is small in terms of number of machines; absent being allowed to sell them to China, there are only a handful of customers for ASML’s product. Setting up a competing source for these machines would be very expensive, so the users of these tools instead have made a choice to allow themselves to be sole-sourced, rather than paying to set up a competitor. 

From a policy perspective, having this expertise exist only in one firm is quite dangerous, and Western governments should consider whether this is a situation that should be allowed to continue.

TSMC is also a natural monopoly. This is not because of the cost associated with bringing up a new process: Spending tens of billions is well within the capabilities of almost any large state, and certainly of many large companies. 

Fabless semiconductor companies that use TSMC’s fabrication services get to tap into a very rich ecosystem of pre-designed IP blocks, from an entire network of different vendors. As a result, TSMC users can often license 90% or more of the design of their chip, while their engineers design only the pieces that generate significant technical differentiation.

Doing so greatly reduces risk, and improves time-to-market. This network effect has not yet been replicated at Samsung or Intel; IP providers flock to the processes where they can capture the most users, and users flock to the processes where they can get access to the most robust basket of licensable IP. 

Because TSMC is so dominant in terms of number of users, the IP vendors use scarce resources to develop for TSMC, instead of Samsung and Intel.

If the US wants to create an onshore alternative to TSMC to counter Chinese threats to Taiwan, building a factory is not enough: The entire ecosystem would need to be duplicated.  

Creating redundancy for TSMC would not be just a matter of standing up processes and factories; it’s a matter of replicating this entire design ecosystem on top of those processes, plus the downstream offerings from TSMC’s partners in the packaging and test ecosystem.  Network effects make building an onshore alternative to TSMC far more difficult and expensive than it might initially appear.

Having a parallel ecosystem in the US is both a national-security imperative, and, ironically, a great way to reduce the likelihood of an attack on Taiwan. 

Right now, the US needs to defend Taiwan for multiple reasons: Taiwan is an ally that shares liberal democratic values, it is the most important source of advanced semiconductor chips, and it is the linchpin of the First Island Chain, helping to bottle up the PLA (People’s Liberation Army) Navy in home waters. 

Removing the dramatic level of US dependence on Taiwanese chips would eliminate one of the incentives for an attack by the People’s Republic.

So long as TSMC remains a Taiwan-based company, the entire chip-manufacturing ecosystem is vulnerable to disruption – even if TSMC builds some factories in the US, as it is doing. The Communist Party of China may one day decide that reunification is a more important consideration than preserving the TSMC foundries.  

And with China rapidly losing access to Taiwanese-manufactured chips because of US export controls and their effect on TSMC, it is entirely possible that the CPC will calculate that destroying TSMC in the short term is a net benefit for China, giving it a comparative advantage.

The China challenge

US semiconductor companies and those located in allied countries must never be allowed to fall into a situation where any critical part of their business – whether it’s things they need to buy, people they need to hire, capital they need to employ, or indispensable customers – flows through adversarial regimes. 

Any dependency on adversaries – and the primary adversary is China, along with its allies Russia and Iran – needs to be ruthlessly and rapidly addressed through aggressive policy measures. This is true not just for chip companies, but for the companies making the inputs to the chip ecosystem, and the companies purchasing the chips. 

Sadly, since the end of the Cold War, these measures have been woefully lacking, and China has been allowed to strong-arm US and allied companies into ever-more-close interactions. With every such interaction, more value and technical knowledge is transferred from the West to China.

The CPC has deployed a full panoply of strategies to advance their domestic semiconductor industry. The CPC has repeatedly demonstrated a willingness to flout treaty obligations, human-rights standards, environmental standards, and all the aspects of the international rules-based order in their pursuit of power.

One strategy that has been used again and again, and which Miller highlights, is the creation of what appear from the outside to be corporate entities, such as Huawei and ZTE. These entities are, in reality, extensions of the Chinese government. 

Their agenda is indistinguishable from that of the state, their ownership is obscure at best, and they operate with massive subsidies from the government which allow them to break into key and strategic industries where they would be unable to compete on economic terms. They leverage the sharp power of the state for espionage, negotiation, and gray-zone warfare against Western companies and their personnel.

Furthermore, these “companies” benefit from an asymmetric regulatory regime. Foreign companies are permitted to sell into the China market only to the extent that they team with local firms in ways that involve mass transfers of intellectual property, while the PRC national champions have, by default, exclusive access to the local market in China.

These national champions are treated by Western governments, by and large, as economic competitors in Western markets, but they get all the advantages of being arms of the state for purposes of developing strategic and key goods for both export and domestic use.

Because of the enormous size of the China market, this asymmetric access offers decisive advantages, and has allowed Chinese companies systematically and successfully to attack a number of markets – including telecom equipment, electronics assembly, the large-scale manufacture of circuit boards, display technology, and many more. 

The CPC is deeply committed to replicating this model for advanced semiconductors, and has already had considerable success in doing so for the less advanced variety.

This agenda is no secret. President Xi Jinping, in addressing the importance of core technologies – specifically semiconductors – is quoted in Chip War (page 248)as saying before an elite audience of Chinese entrepreneurs and technologists:

“We must promote strong alliances and attack strategic passes in a coordinated manner.  We must assault the fortifications of core technology research and development…. We must not only call forth the assault, we must sound the call for assembly, which means that we must concentrate the most powerful forces to act together, compose shock brigades and special forces to storm the passes.”

According to Chris Miller, Xi’s audience “counted on” the use of non-economic means when seeking a competitive commercial advantage.

One interesting point that Miller makes is the high cost of building advanced fabs, and (as a separate example) of replicating the EUV machines being built by ASML. As such, he is somewhat sanguine about the danger of China duplicating these advanced capabilities; he seems to believe that the high cost will prevent them from catching up.

But Miller misses a crucial point: Building a cutting-edge semiconductor foundry process costs several tens of billions of dollars only under a very particular set of circumstances. Access to the Western ecosystem of suppliers is required. For it to cost tens of billions, the entity setting up the foundry needs to be able to buy the tools and chemicals and wafers and masks and design software at market prices. 

If, on the other hand, access to all of those items is denied, then the cost to build a cutting-edge foundry would be in the trillions of dollars, since it would be necessary to replicate the entire ecosystem. 

Tens of billions of dollars is a check that the CPC would be delighted to cut for advanced foundry capabilities. Trillions is not feasible. This difference is critical. Allowing China access to the Western semiconductor ecosystem cuts the cost of China standing up an independent ecosystem by a factor of perhaps 100, and makes doing so feasible.

With the China advancing autocracy and threatening its peaceful, democratic neighbors, there is no obvious policy justification for the liberal democratic regimes of the world continuing to allow China any access whatsoever to this ecosystem.

The Western response

How should the West respond to the China challenge? Miller’s book largely eschews policy recommendations, but it is easy to derive a number of lessons from the stories in his book and from our experiences in the worlds of semiconductors and geopolitics.

In the US, the ongoing debate about how to help the semiconductor industry tends toward two orthogonal schools of thought. 

The first are free-traders, who believe that the most important policy questions revolve around how we can make the semiconductor and high-tech industries in the United States maximally competitive and efficient. They tend to believe that government intervention in the semiconductor industry will only create structural inefficiencies, because the people who know how to pick economic winners tend not to go into government. 

They believe that less regulation will make the industry more efficient and more competitive. And in a world where chips are primarily an economic good, and our companies can compete primarily on an economic basis, this argument is sound and powerful.

The second group are those who see semiconductors primarily as a “key” or a strategic good, and who believe that strong government intervention is needed in order to ensure the long-term competitiveness of our on-shore semiconductor industry. They tend to focus on the massive subsidies and the use of state power to bolster Chinese competitors vis-à-vis Western firms. They argue that similar measures are needed in the US.

There is a third path, which is more nuanced and gets less attention, that preserves the distinctive advantages of America’s capitalist, open system for producing and fostering innovation, and allows for competition and specialization not just among firms but also among allied states. 

This third path is to resurrect the Western entente of the Cold War, not just as a military alliance but also as an exclusive trade zone for high-technology goods. Free trade of high-tech goods with free people, those who share US values and interests, ought to inform the economic policies of the liberal-democratic alliance.

China, Russia, Iran, North Korea and their allies and clients should no longer be allowed access to America’s high-tech goods and services, and should be comprehensively frozen out of access to the entire semiconductor ecosystem. This most especially should include the tools to build and design semiconductor chips, the raw materials and fine chemicals needed as inputs to the factories, and the spares required for their existing factories.

The price of participation in the Western ecosystem for US allies will necessarily be a strong commitment to preventing their citizens and companies from working with PRC-controlled and allied entities in the semiconductor space. 

This same regime will need to be expanded over time to other industries producing “key” and strategic goods; only by doing this will it be possible to preserve the distinct advantages of an open market among the free peoples of the world.

Companies – especially smaller ones – are generally not in any position to hedge unforeseen geopolitical risks. They go where it’s cheap to design, build and package chips, within the constraints of law – both local laws, and those of the states where they have their headquarters and primary operations.

When geopolitical realities change, it can and does create enormous economic disruption.  Requiring Western semiconductor companies to bear the costs of decoupling, and endure the sudden loss of a large fraction of the market for their goods, would be enormously damaging. 

The cost of keeping these companies in business and allowing them time to restructure their businesses is part of the cost of decoupling from China, and is not corporate welfare.

This is especially important in a semiconductor downturn, like the one that is starting now.  Semiconductors are a highly cyclical industry, and making sanctions against China stick is much easier when there is strong demand for capacity, in a hot market, than in a downturn. 

In down-cycles, contracts from firms in the PRC may be the only thing keeping a supplier of tools or chips afloat. Trying to freeze PRC firms out of the Western ecosystem at the beginning of a major semiconductor down cycle introduces the risk of irreparable short term harm to crucial Western firms. 

Structuring short-term subsidies to buffer the lost revenue from rapid regulatory changes is necessary to make sure that the US doesn’t lose these firms. To preserve the competitive nature of the Western semiconductor ecosystem, any subsidies to support the cost of decoupling must be a temporary bridge, and not a long-term crutch.

Without these kinds of policy changes, there will be enormous political and lobbying pressure from semiconductor companies that sell to China, demanding exceptions and waivers from any action that drives rapid decoupling. With current regulatory strategies, companies are already responding both with political pressure and by developing products that skirt the edges of the newly introduced policies.

What exceptions should the United States allow within such a policy? Only the ones that create long-term dependency and insecurity for its adversaries. Only chips that are several generations behind the cutting edge should be sold to adversary regimes and their aligned companies, and only in the form of finished systems, rather than as bare chips.

The US and its allies should, on security grounds alone, entirely forbid the import of chips or associated systems that are made, packaged or designed in China. It should sell only end-user systems that are generations back from the state of the art, and which will cause the CPC leadership to question whether the systems on which they depend can be relied upon in a crisis.

Conclusion

China is at war with the West. This war is being fought today primarily through economic means – not because of any qualms among the leadership of the CPC about using kinetic measures, but because they believe that economic warfare is better suited to their ends at this time.

The goal of the CPC is for China to become the essential power, and to exert hegemony over the rest of the world. Any liberal democratic regime that can deny key or strategic goods to China is a threat to that agenda. 

Right now, China is engaged in a long-term effort to subsume the independence of the liberal-democratic states of the world by acting outside the strictures of the rules and free-trade based international order that the United States championed in the aftermath of the Cold War.

Therefore, the US should move expeditiously to reinforce its relationships with liberal-democratic allies; allowing them to compete for government subsidies, as part of an exchange for eschewing work with PRC entities, is very desirable and was an essential part of reinforcing America’s Cold War alliances. 

The recognition and reinforcement of this geopolitical reality will reinforce a liberal-democratic alliance of free peoples, while reducing authoritarian access to this entire ecosystem.

As Adam Smith noted in The Wealth of Nations, some goods and sectors enhance security, not just opulence. The US and its allies need to take a careful measure of which goods are purely economic, which are strategic, and which are key. 

The US needs to construct a system where it can engage in open, capitalist competition with its allies and their corporations. And it needs to stop letting its adversaries mine America’s key and strategic industries for their own benefit, which has been to the detriment of the free peoples of the world.

Source

Every nation-state faces a fundamental choice: With whom will we engage in trade?

What goods and services will we sell, and to whom? What will we buy, and from whom? For authoritarian regimes, where the state subsumes commercial activities, the answer is simple: Only transactions that enhance the power of the state are permitted. 

While this principle may be suspended temporarily (for example the New Economic Policy in Russia, or the temporary Chinese market reforms of the Deng-Jiang-Hu era), or may be incompetently implemented, authoritarian regimes exercise full control over trade, in order to enhance state power.

For commercial republics, by contrast, these choices present thorny dilemmas; commercial republics generate massive wealth from free trade. As a result, it is often difficult to differentiate between commercial and national interests; more trade generates more wealth, which in turn generates more power and security.

Except when it doesn’t.

The semiconductor industry, as the book Chip War by Chris Miller, assistant professor of international history at the Fletcher School of Law and Diplomacy, vividly illustrates, is one where free trade has provided enormous benefits to the national security of the United States and its allies. 

This played out first during the Cold War, both through the economic and societal impact of the digital revolution and through the many direct impacts on military systems and capabilities. The wide-reaching implications of Moore’s law in the United States provided a crucial technological edge over our adversaries. 

The emergence of precision targeting and smart weapons systems offset huge numbers of Soviet weapons with far fewer of America’s own. The drive to increase the complexity and decrease the cost of semiconductor devices involved national specialization, including both chip-building collaborations with Japan, Taiwan and South Korea, and partnerships throughout Southeast Asia and China focusing on packaging and testing these chips and assembling them into finished products. 

These relationships strengthened ties among the alliance that the United States forged to contain the Soviet Union.

Although the last part of the book focuses on the emergence of an adversarial relationship with China around chips, a key distinction that is implied and illustrated, but never quite explicitly articulated, is between these international collaborations among friendly nations and the more recent interactions with China. 

During the Cold War, semiconductor companies were free to collaborate with companies and nationally sponsored entities across the free world, but were not permitted to interact with the state-sponsored semiconductor development efforts within the Soviet Union. Selling semiconductors and transferring associated technology to China was in the national interest, because – in the aftermath of Richard Nixon’s and Henry Kissinger’s initiative in 1972 – China was part of the alliance against the Soviet Union.

But at the end of the Cold War, American policymakers mistook the benefits of national specialization and decided that anything that increased world trade must be an unalloyed good for the United States. 

Even when free trade came at the cost of sending the secrets to building our technological “crown jewels” to China, the justification was that commercial engagement would produce wealth, wealth would result in a larger Chinese middle class, and, inevitably, a Chinese middle class would demand political liberalization and reform.  This was wishful thinking of the worst kind.

The Communist Party of China (CPC) took full advantage of American naïveté, capturing a near-monopoly in the lower-value portions of the semiconductor assembly chain – most notably in consumer-scale board and system-level assembly. China has, however, struggled mightily to design and build the chips itself; the chips are the highest-value and highest-complexity components of nearly every modern consumer and military product. 

China’s efforts have resulted in onshore capabilities that are now only about two generations (that is, about five years) back from the cutting edge at Intel, Samsung and Taiwan Semiconductor Manufacturing Company; 15 years ago they were much further behind. In 2008, when TSMC was making 40-nanometer transistors, China’s Semiconductor Manufacturing International Corporation (SMIC) – its national champion in this space – was at 300nm. This represents a five-to-six-generation difference.

If you want historical context for how this happened, and to understand the key personalities, decisions and incentives that drove what might seem from the outside to be a set of deeply perverse outcomes, Chip War is a great place to start. The book is, at its core, a history of the geopolitics of semiconductors.

Chip War: the book

Chip War achieves something that few books about the semiconductor industry manage: It provides enough context for those outside the field to understand the nuance and complexity of this important industry without drowning the reader in technical jargon and detail. 

The book does this by telling the story of the semiconductor industry, more or less from its inception, through a series of short vignettes. Each vignette addresses a particular person and episode, moving chronologically through the industry’s development. 

This is a lot of ground to cover, and it becomes clear that Chris Miller must have selected the most telling and entertaining anecdotes out of perhaps thousands gathered from conversations with leaders of the semiconductor industry.

In Chip War, Miller focuses primarily on the most advanced digital semiconductor chips; as a result there are entire swaths of the industry that are barely mentioned.

It would be possible for someone outside the semiconductor industry to get the impression that, at a geopolitical and geostrategic level, only cutting-edge processors and memory chips matter. This is far from the truth, since a chip without a package, or a packaged chip without a board to carry it, is useless. 

And the semiconductor world is much more diverse than just the advanced digital chips on which this volume focuses. It is both deeper vertically, in terms of the tools, intellectual property (IP) and software needed to design, build, package and test these chips, and broader horizontally, in terms of the many other kinds of chips made in silicon and other materials.

Miller emphasizes the “bewildering” and “dizzying” complexity of semiconductor supply chains and international interconnections in the chip industry. And indeed, the products that the chip industry produces are incredibly complex – the most complex products, by many measures, built by humans. 

The physics and engineering that go into building microchips is subtle, and requires extremely deep, specialized knowledge. But the industry itself is no more complex than any other large, highly cyclical, fast-growing technical industry. Conflating the complexity of the product with the complexity of the industry is a mistake.

Semiconductor companies are no more complex, as businesses, than any other. Think of any semiconductor industry company as a box that has inputs, outputs, and assets. 

The inputs are capital, talent, tools, and the items that go onto the bill of materials, which is the list of input materials and components that the company transforms into their product.  The outputs are IP (in the form of patents or licenses), products, trained people and (it is hoped) profits. 

And then there are assets. These can be trade secrets, equipment, human capital and – often most important – a corporate culture that is well-adapted to efficiently converting the inputs into the outputs. Pretty much any company can be modeled this way, independent of the complexity of their actual product.

By thinking about the semiconductor industry in these simplified terms, it becomes possible to look at the interactions both among companies, and between companies and governments, in a systematic way. 

One of the things that becomes clear from the vignettes in Chip War is that, even when chip companies have tried to remain largely independent of government influence, they’ve found that their interactions with governments – both friendly and hostile – are of existential importance. Anywhere that a government can touch any of the inputs, outputs, or assets of a chip company, that government has an opportunity to gain leverage. And the governments of the world have not been hesitant to use that leverage.

One of the benefits of telling the story of this industry through vignettes is that the importance of individual incentives and individual decisions comes through very clearly. The dynamic relationship between macroeconomic incentives and microeconomic behavior of the firms and leaders plays out again and again through the pages of Chip War.

Nation-states and the chip industry

The trajectory of the semiconductor industry depended in part on the interaction of that industry with various nation-states.

Since the formative days of the semiconductor industry in the United States, one dynamic was the high rate of exponential growth. This growth arose from rapid increases in committed capital, number of people involved, and consumer impact, all running in parallel. 

With exponential growth, a very small industry can grow to be a significant chunk of the GDP seemingly overnight. Furthermore, some of the most important decisions revolved around how these firms chose to interact with the US government, and which governments they chose to work with.

Chip War makes it clear that, from the very beginning of this industry, the companies at the leading edge of silicon design and chip fabrication technology were subject to pressure to move their technology offshore (first to Japan, and later to China and elsewhere).

At the same time, the adversaries of the United States were devoting enormous resources to replicating the products of the semiconductor industry. 

The Soviet Union was successful in maintaining their semiconductor capabilities at about two to three generations (that is, about five years) behind that of the West for almost the entirety of the Cold War. It did this largely through espionage and appropriation of Western intellectual property, rather than through internal innovation. 

Because the Soviets did not have access to Western markets, the flow of information, knowledge, equipment, software, and personnel was dramatically limited. Furthermore, getting chips with billions of transistors to work requires a lot of data and very accurate statistics, and that requires building enormous numbers of chips. 

The Soviets chose to be fast-followers and to only use silicon chips where they were absolutely required, rather than making them a pervasive pillar of their economy. As a result, they were never able to catch up.

Getting to massive scale has always been a key advantage in this industry; building more copies of the same chip makes it easier to collect the statistics needed to improve performance and yield, thus enabling higher and higher complexity for each subsequent generation. This dynamic has contributed to a dramatic concentration of the semiconductor fabrication business over time, into just a few giant players.

One of the most dramatic stories the book relates is about how Taiwan rose to become the chip powerhouse that it is today. 

The ingredients were simple: First, a brilliant engineer and businessman, Morris Chang, who had spent decades becoming the master of his trade in the United States before being persuaded to move to Taiwan to build up its semiconductor industry. Second, strong backing from the government, across decades, to support all of the needs of what became their most important company, TSMC.

Children in Taiwanese elementary schools today learn what a wafer stepper does and can explain how semiconductor wafers are made. Taiwan has engaged in a whole-of-society effort to own a big, critical piece of the semiconductor ecosystem. This has been wildly successful because, as the author points out, the Taiwanese have engaged very closely with Silicon Valley and the rest of the Western ecosystem.

Taiwan concentrated its efforts on a few very specialized niches where government intervention could be extremely effective, in support of aggressive commercial efforts. TSMC is, in effect, a direct result of business model innovation, relentless technical and commercial execution, and extremely effective industrial policy.

Chips: economic, strategic, and key products

Why is it so difficult for the leading companies to function without government intervention in this industry? In short, because the semiconductor industry is so important.  Semiconductors are simultaneously an economic good, a strategic good, and a “key” good.

As an economic good, the chips are subject to market forces. The companies that make and sell them compete with one another on price, reputation, quality, execution, time-to-market and performance. Cutting-edge digital chips, which require the most expensive fabrication processes and design tools, tend to differentiate more on power consumption and performance, while trailing-edge ones tend to differentiate more often on price (though these are broad generalizations).

One thing that Miller conveys very clearly is that semiconductor chips are the digital “brains” behind every single “smart” product in the world. Any device that processes information, from toasters to cellphones, to cars and airplanes, remote controls to modern lightbulbs, depends on semiconductor chips. 

If it has a display, or communicates by Wi-Fi or Bluetooth, or even if it just uses electricity, there are almost certainly silicon chips somewhere inside making it work. These chips can be as simple as power regulators or converters, or as complex as cutting-edge digital processors.

As a strategic good, semiconductor chips are the raw material of warfighting. Every modern military system depends on chips to function. Satellites, radar systems, missiles, tanks, planes, Global Positioning System (GPS) receivers, inertial and optical guidance systems – all of it relies on semiconductor chips. 

Military radios rely on chips for encoding and decoding signals. The red-dot sights on rifles require LEDs, which are small chips that produce light. Rangefinders require lasers and chips for computation. Logistics are managed through computerized databases, and parts are tracked with barcode scanners and RFID (radio-frequency identification) systems, all of which rely on silicon circuits. Chips are the raw material for nearly every aspect of modern warfighting.

And furthermore, chips define the arena in which gray zone wars are fought. Any form of digital warfare, whether it is direct – that is, hacking an adversary’s infrastructure and systems (such as Stuxnet) – or indirect, via an influence operation through a social media platform. 

As artificial-intelligence (AI) systems become more capable, conflicts between adversarial autonomous systems will become more common: The side with the better hardware will have a huge advantage in such conflicts, because their systems will be smarter and faster.

From a security perspective, relying on chips designed or fabricated by an adversary regime is a most uncomfortable position. Setting aside the obvious challenges that emerge from losing access to replacement hardware, there is a second, more insidious problem: Digital chips have become so complex that there is no way to know what they contain. 

Unless you control the design and trust the fabrication facility, there is no way to know whether something has been inserted into the chip that creates a hardware vulnerability or a backdoor for an adversary. Detecting such an insertion as an end user is nearly impossible.

Regardless of what software is run on top of untrustworthy hardware, there is no real security. Furthermore, adding software layers to try to create some security on top of untrustworthy hardware will necessarily degrade performance. 

In a modern conflict, the default winner is the one who designed and built the hardware and software that their adversary is using. This undoubtedly creates considerable uncertainty and discomfort in China today, given its near-100% dependence on the West (including Taiwan) for advanced semiconductor chips.

And as a “key” good – to deploy the term Halford Mackinder used in his Democratic Ideals and Reality – semiconductor chips “unlock” entire swaths of the economy. Without them, entire sectors of the economy would rapidly grind to a halt. 

Cellphones, for example, have an average lifespan of just a few years; they have to be frequently replaced. We’ve recently been able to observe the dramatic consequences of losing access to chips in the automotive industry: For the lack of a few dollars’ worth of chips, it is impossible to complete the manufacture of a US$50,000 or $100,000 vehicle. 

The same is true across multiple domains: Without cutting-edge central processing units (CPUs) and graphics processing units (GPUs) from AMD, Intel and Nvidia, the rollout of new services from Facebook, Amazon and other hyperscalers grinds to a halt. 

Even in cases like Apple and Google, which design many of their own chips, access to the chip fabs at either TSMC or Samsung are essential to their businesses. Without comparatively inexpensive chips that might cost tens or hundreds of dollars, airplanes cannot fly, trains cannot move, and satellites cannot be launched.

The chips are a “key” good in each of these industries. Other parts can be replaced, or at least built in multiple places – the chips can, in many cases, only be built at one factory.

And very often that factory is a TSMC fabrication facility, located in Taiwan. Right now, TSMC is the default option for fabricating cutting-edge digital chips, because it is the best foundry in the world. 

As a foundry, TSMC’s business model is to provide the capability to build extremely complex chips to other companies. With a design team and a pile of money, it is possible for any small company to design a custom silicon chip, and TSMC will make it and deliver it.

It is more informative to think of TSMC as a company that delivers the world’s most advanced processes for building semiconductors in volume, at scale, than it is to think of them as a company that builds chips. 

TSMC’s core mission is to bring a new, more advanced fabrication process online every couple of years, defining the state of the art for building advanced digital chips. These fabs are operated as foundries, as a service for the world’s chip design companies – these design companies are known as “fabless” semiconductor companies.

Samsung competes, especially for giant customers, but TSMC gets the lion’s share of the business. Intel has historically been an “integrated device manufacturer,” running chip fabrication facilities solely to build its own products. 

Intel is currently retooling its business to act as a foundry, but the transition from fab to foundry is likely to be extremely challenging. A foundry, while it uses very similar facilities to an integrated device manufacturer, is a fundamentally different business. Making the jump into a foundry business has historically been a long and painful process.

TSMC, Samsung and Intel make cutting-edge digital chips. These are the chips that lead Moore’s law, getting ever denser and more complex, with billions of transistors in a single chip today. Such chips require fabrication plants that can produce devices that are only a few nanometers wide. 

Such fabs are hugely expensive – tens of billions of dollars for a single factory, and far more for the full R&D effort to bring up a process.

While most of Chip War is focused on the design and manufacturing ecosystem for the world’s most complex, advanced, digital semiconductors and memory chips, this is far from the only part of the industry that produces key, strategic goods of incredibly economic importance. Another area that deserves attention is the trailing edge.

Many other technologies, often lumped into “more than Moore” as a catch-all term, are in fact implemented in processes that don’t need few-nanometer linewidths or involve billions of transistors. 

These chips are made in semiconductor fabs, and participate in the same ecosystem of tools, software, test equipment, chemicals, etc. Entire industries rely on these devices – light-emitting diodes (LEDs), lasers, microelectromechanical systems for position sensing, optical detectors, fluidics for health care and diagnostics, silicon photonics, high-power analog, high-speed and high-power-radio-frequency integrated circuits, radiation-hardened electronics and many others – that are made solely or primarily in trailing-edge foundries. 

These foundries compete not by making linewidths narrower, but through process innovation, integrating novel materials, and making devices other than transistors.

These activities often advance faster than Moore’s law in their own critical performance metrics, generating goods that are simultaneously economic, strategic and “key,” but they are built in foundries that are orders of magnitude less expensive than those at the cutting edge of digital transistor fabrication. 

They don’t need the ultra-fine linewidths of cutting-edge CMOS (complementary metal-oxide-semiconductor) electronics or memory circuits, making the fabs much more affordable. 

These “trailing edge” fabs are far more important than the name implies: While their equipment may be trailing-edge for making digital transistors, these fabs build cutting-edge products for applications beyond digital, and their monumental importance is widely under-appreciated, especially in policy circles.

Intel’s chief executive officer, Patrick Gelsinger, has recently been explicit about the geo-strategic implications of semiconductor manufacturing, suggesting, “Where the oil reserves are defined geopolitics for the last five decades. Where the [chip] fabs are for the next five decades is more important.”

Chokepoints in the chip ecosystem

Policy discussions about semiconductor issues often neglect to incorporate the fact that chips are not interchangeable. From an end-user perspective, a CPU chip from AMD or Apple provides the same functionality as one from Intel; they all support a software ecosystem that allows the end user to run Word or Gmail on their computer. 

But these processors are not interchangeable; far from it. A computer that uses a chip from AMD requires different boards and supporting chips than the one for an Intel chip. Only at the system level can these chips be used interchangeably, and that system-level interchangeability is only true for a small subset of the systems that rely on chips.

This same dynamic is true across the ecosystem. While Cadence and Synopsys and Mentor Graphics all make software that does very similar things, these tools are not easily interchangeable. A chip design team that has made a multi-year investment in a Cadence-based design flow would lose years of effort transitioning to one from Synopsys, or vice versa.

The same is true for IP blocks, which are completed, tested, field-proven circuits that can be included in a chip design as a pre-completed block: Switching a circuit design from one IP block to another is often a nearly complete redesign. 

For instance, ARM and RISC-V architectures both have commercially licensable IP blocks for CPUs that can be included in a chip design. While both blocks perform fundamentally the same function, replacing one with the other requires enormously disruptive changes to both hardware design and to the software and firmware that live higher in the engineering stack. Switching a product from one to the other could take multiple years and cost hundreds of millions of dollars.

Because of these dynamics, there are a lot of chokepoints in the semiconductor ecosystem, where, if a specific good or service is removed or disrupted, there is an amplifying cascade of downstream effects. 

For instance, Miller writes in detail about how the entire advanced semiconductor ecosystem is dependent on the multi-hundred-million-dollar EUV (extreme ultraviolet) lithography systems from ASML, which exercises a monopoly on these tools. 

This industry is rife with extremely sticky suppliers, where the cost in time and effort to switch from one to another is measured in years. Any disruption to the supply of machines, spare parts, software, design IP, masks, wafers, etc from a single supplier can disrupt an entire swath of the industry, jeopardizing the viability of multiple companies.

It gets worse. Even when there are multiple sources for standard commodity semiconductor chips, with a standard physical and logical interface, it is not always possible to use the parts interchangeably without a considerable engineering and qualification effort. 

Because the specifications for chips are complex, even nominally identical products are often not perfect replacements for one another; the only way to know is to build systems with both sources of parts, and put them through a full reliability and qualification cycle. Doing so can easily take more than a year, and that’s if the tests get passed the first time, and no design changes are required to accommodate the alternative source of chips. 

Large companies that are producing a product in volume will go to the trouble to qualify multiple sources for critical parts (when they can) in order to make their supply chains more robust. But this is utterly impractical for smaller companies, and it’s an enormous logistical and technical burden to qualify multiple suppliers for more than a tiny fraction of the chips that go into any complex system.

Another strategy to increase supply-chain robustness is to hold a buffer stock of critical components. But this is extremely expensive, and can have a dramatic impact on the economic competitiveness of the products. Fat supply chains are especially expensive in an industry where Moore’s law makes cutting-edge chips, and the systems built with them, obsolete every couple of years.

Ironically, the behavior required by the key and strategic nature of semiconductor products drives the industry toward developing robust, buffered supply chains, while the imperatives associated with economic goods drive the industry toward lean manufacturing and just-in-time delivery. 

During times when geopolitical drama is absent, the companies have to run lean to survive. But when national interests come into conflict, lean supply chains collapse, and national survival demands fat, redundant supply chains. This represents a fundamental policy challenge.

This is especially true when we take into account the fact that economic survival is in and of itself a strategic good. A country whose economy collapses because of lack of a key good is in no shape to successfully prosecute a war. We’re seeing the beginning of this dynamic in Russia, as it loses access to Western chips and other high-tech goods.

Corporate behavior is shaped not only by supply chains, but also by demand, talent and capital chains. 

One thing that comes through both in Miller’s book, and in decades of personal experience in this industry, is that autocratic governments have a rich history of seeking to insert themselves into these chains in any way that they can – especially at chokepoints – in order to exert leverage, exfiltrate IP, stand up their own chip design and manufacturing capabilities, and shape corporate behavior to their ends.

This is the first article in a two-part review. For Part 2, click here.

Source

SpaceX has just unveiled its Starshield satellite network service intended for government and military use, just as the US military has conducted tests with the company’s Starlink low-Earth orbit (LEO) satellite constellation. 

This week, SpaceX revealed its Starshield satellite project on its website, advertised as leveraging Starlink’s technology and launch systems for government use, focusing on earth observation, communications, and hosted payloads – which means clients can load custom equipment in the satellites. 

Starshield is touted as offering unparalleled end-to-end user data encryption, with the high-assurance cryptographic capability to host classified payloads and securely process data. It also uses Starlink’s inter-satellite laser communications terminal, allowing it to integrate other satellites into its network. 

It is also claimed to have an end-to-end logistics chain, with SpaceX providing all the hardware from ground terminals, antennas, rockets and satellites and operating the entire constellation. This arrangement may address US supply-chain issues wherein manufacturing of sensitive electronics is outsourced to China, giving the latter espionage opportunities on classified information or manipulating these systems entirely. 

While there is minimal information about Starshield, its unveiling comes after the US military tested Starlink. Last month, Bloomberg reported that the US had tested Starlink as a possible way to maintain communications with remote military outposts in the Arctic, an increasingly contested region with China and Russia.

The timing of Starshield’s unveiling may indicate that, at present, Starlink needs military-specific capabilities, which Starshield aims to provide.

The military value of LEO satellite constellations has long been understood yet has only been recently validated. 

In a 2004 RAND research brief, Robert Preston wrote about the advantages distributed satellite constellations have over monolithic constellations focused on a few powerful satellites. He said distributed LEO satellite constellations cost less to launch, perform better during deployments, are more failure-resistant, and are more survivable in an attack. 

In line with that, the Ukraine war may have validated the military value of LEO satellite constellations, vividly demonstrating this space-based capability’s tactical and strategic effects. 

In March, a report by the French Institute of International Relations noted the military advantages satellite constellations such as Starshield can bring, such as shorter development time and launch economy of scale, increased performance from proliferation, and exploiting dual-use features of satellite technology. 

The report also says that soon, real-time satellite imagery can be requested and received by ground commanders to support decision-making. However, this capability remains only in the hands of the world’s leading military powers.

The Ukraine war has spurred a race by major military powers to acquire satellite constellations, with private companies taking the lead in offering this capability to governments and militaries. 

In April, the South China Morning Post reported that private corporations are lining up to provide space-based satellite capabilities to paying government and military clients in the West from Germany, the US and Canada. Companies such as Airbus Defense and Space in Germany, Telesat in Canada, and Kuiper and Starlink in the US are at the forefront of the commercialization of this increasingly vital military capability.

However, China may already be a step ahead of the US in having its dedicated military LEO satellite constellation, as it has been operating the Yaogan LEO satellite constellation since 2006, in contrast to Starlink, which SpaceX first launched in 2019. 

Global Security describes the Yaogan as primarily used for ocean surveillance along China’s littorals, providing real-time command and control intelligence support to commanders afloat and ashore, and receiving, processing, and disseminating timely all-source information on mobile and stationary ocean and ground targets. 

Further, in a June 2021 blog entry, Steve Hsu notes that China has been using satellites to track US warships in the Pacific. In addition, identical satellites can provide targeting data for their anti-ship ballistic missiles. 

Moreover, in a separate blog post from November 2021, Hsu notes the vulnerability of large targets such as aircraft carriers, saying they are easy to track with satellites and that advances in artificial intelligence (AI), machine learning, and sensor technology favor the attacker, not the defender.

He notes that advances in those fields may now have enabled China to hit US carriers with greater accuracy, needing its Yaogan satellite constellation to provide the location of the target to within tens of kilometers and let its missiles’ onboard sensors and AI guide them in their terminal stage to the target. 

Despite that, Hsu says it might not be necessary for China to sink a US carrier to serve as a warning shot, as China now has the satellite capability to track all container ships and tankers approaching Kaohsiung in Taiwan or Nagoya in Japan. So all that China needs to do, Hsu says, is to sink one container ship or oil tanker or issue a solid warning to deter all civilian maritime traffic from approaching Taiwan. 

Aside from possessing military satellite constellations, China is also taking a page from the West in commercializing LEO satellite services. In January, the South China Morning Post reported that China was establishing a rival to the Starlink satellite constellation. The source said Beijing-based startup GalaxySpace aimed to establish an LEO constellation of 1,000 satellites to compete with Starlink in providing high-speed Internet in remote areas. 

Although smaller than Starlink, the source notes that it will be the first to use fifth-generation (5G) technology and that its customers will likely be overseas companies, the Chinese government, and the military. Thus GalaxySpace may be China’s project to provide the Internet as an emerging public good for the international system, reinforcing its status and legitimacy as an emerging power.

Source

If you're one of the billions of people who have posted pictures of themselves on social media over the past decade, it may be time to rethink that behavior. New AI image-generation technology allows anyone to save a handful of photos (or video frames) of you, then train AI to create realistic fake photos that show you doing embarrassing or illegal things. Not everyone may be at risk, but everyone should know about it.

Photographs have always been subject to falsifications—first in darkrooms with scissors and paste and then via Adobe Photoshop through pixels. But it took a great deal of skill to pull off convincingly. Today, creating convincing photorealistic fakes has become almost trivial.

Once an AI model learns how to render someone, their image becomes a software plaything. The AI can create images of them in infinite quantities. And the AI model can be shared, allowing other people to create images of that person as well.

John: A social media case study

When we started writing this article, we asked a brave volunteer if we could use their social media images to attempt to train an AI model to create fakes. They agreed, but the results were too convincing, and the reputational risk proved too great. So instead, we used AI to create a set of seven simulated social media photos of a fictitious person we'll call "John." That way, we can safely show you the results. For now, let's pretend John is a real guy. The outcome is exactly the same, as you'll see below.

 

In our pretend scenario, "John" is an elementary school teacher. Like many of us, over the past 12 years, John has posted photos of himself on Facebook at his job, relaxing at home, or while going places.

 

Enlarge / These inoffensive, social-media-style images of "John" were used as the training data that our AI used to put him in more compromising positions.

Ars Technica

 

Using nothing but those seven images, someone could train AI to generate images that make it seem like John has a secret life. For example, he might like to take nude selfies in his classroom. At night, John might go to bars dressed like a clown. On weekends, he could be part of an extremist paramilitary group. And maybe he served prison time for an illegal drug charge but has hidden that from his employer.

 

At night, "John" dresses like a clown and goes to bars.

 Ars Technica

 

"John" beside a nude woman in an office. He is married, and that's not his wife.

 Ars Technica

 

"John" spends time on weekends training as part of a paramilitary group.

 Ars Technica

 

John relaxing shirtless in his classroom after school.

 Ars Technica

 

John served time in prison for drug charges just a few years ago and never told the school system.

 Ars Technica

 

John in a great deal of pain, or perhaps doing something else. We've cropped out the operative parts.

 Ars Technica

 

We used an AI image generator called Stable Diffusion (version 1.5) and a technique called Dreambooth to teach AI how to create images of John in any style. While our John is not real, someone could reproduce similar results with five or more images of any person. They could be pulled from a social media account or even taken as still frames from a video.

The training process—teaching the AI how to create images of John—took about an hour and was free thanks to a Google cloud computing service. Once training was complete, generating the images themselves took several hours—not because generating them is slow but because we needed to sort through many imperfect pictures (and use trial-and-error in prompting) to find the best ones. Still, it's dramatically easier than attempting to create a realistic fake of "John" in Photoshop from scratch.

 

Thanks to AI, we can make John appear to commit illegal or immoral acts, such as breaking into a house, using illegal drugs, or taking a nude shower with a student. With add-on AI models optimized for pornography, John can be a porn star, and that capability can even veer into CSAM territory.

 

We can also generate images of John doing seemingly innocuous things that might still personally be devastating to him—drinking at a bar when he's pledged sobriety or spending time somewhere he is not supposed to be.

 

He can also be put into fun and fantastic situations, like being a medieval knight or an astronaut. He can appear young or old, obese or skinny, with or without glasses, or wearing different outfits.

 

John does medieval LARPing once a month.

 Ars Technica

 

John was once an astronaut.

 Ars Technica

 

The synthesized images are not perfect. If you look carefully, a knowledgeable person can spot them as fakes. But the tech that creates these images has been progressing rapidly, and it may soon be completely impossible to tell the difference between a synthesized photo and a real one. Yet even with their deficiencies, any of these fake images could plant devastating doubts about John or potentially ruin his reputation.

 

You can see many examples of people using this same technique (with real people) to create whimsical, artistic profile photos of themselves. And commercial services and apps like Lensa have recently emerged that promise to handle the training for you. What they don't show you is the potential negative effects of this technology if a person uses someone else's face without their consent.

How does it work?

Enlarge / This AI-generated "photo" of ancient Greek philosopher Heraclitus pulls from what the AI model knows about photos of old men and depictions of Ancient Greek culture that were found in its training data set.

Heraclitus of Ephesus

 

If you haven't been paying attention to the rapid progress in AI image generators recently, seeing what we've pulled off above might be very alarming. Basically, computer scientists have figured out how to generate new photorealistic images of anything you can imagine by teaching AI using real photos, and the technology has accelerated rapidly over the past year.

 

The tech has been controversial because, aside from photos, it has also allowed people to generate new artwork that imitates existing artists' work without their permission.

 

One of the most impactful AI image generators is called Stable Diffusion. It's a deep learning image synthesis model (a fancy term for AI software) that can generate completely new images from text descriptions. It can run locally on a Windows or Linux PC with a beefy GPU, on a Mac, or in the cloud on rented computer hardware.

 

With financial support from Stability AI, an academic organization called CompVis trained Stable Diffusion's AI model using hundreds of millions of publicly accessible images downloaded from the Internet. Stability AI released Stable Diffusion as open source software on August 22, 2022, meaning anyone can use it for free, and it has become integrated into a growing number of commercial products.

 

Through intensive training, Stable Diffusion's neural network has learned to associate words and the general statistical association between the positions of pixels in images. As a result, you can give Stable Diffusion a text prompt, such as "Morgan Freeman in a classroom," and you'll get back a completely new image of Morgan Freeman in a classroom.

 

Making images of Morgan Freeman is easy because there are probably hundreds of photos of him in the data set used to train Stable Diffusion. It already knows what Morgan Freeman looks like. But if you want it to make images of an average person like John, you need to give Stable Diffusion some extra help.

 

That's where Dreambooth comes in. Announced on August 30, 2022, by Google researchers, Dreambooth uses a special technique to teach Stable Diffusion's AI model new subjects through a process called "fine tuning."

 

Initially, Dreambooth had nothing to do with Stable Diffusion, and Google held back the Dreambooth code due to concerns of misuse (like we've shown here). But not long after its announcement, someone adapted the Dreambooth technique to work with Stable Diffusion and released the code freely as an open source project.

 

Since then, Dreambooth has become a popular way for AI artists to teach Stable Diffusion new artistic styles (or about products like shoes), and they share their resulting Dreambooth-tuned models in repositories online. Other artists download them and can use previous Dreambooth training to achieve similar effects.

 

Enlarge / Google researchers used cute corgi dogs to illustrate Dreambooth's capabilities to avoid obvious ethical issues about training AI models with humans.

Google

 

But from the beginning, Google's researchers knew that if you can put humans (like John) in any fake situation, trouble might follow. That’s why when they announced Dreambooth, they used pictures of Corgis and not people as examples. Avoiding these potential ethical issues by focusing on images of cute animals has drawn criticism from MIT Technology Review, and it's an example of how AI scientists often try to minimize or avoid criticism for the potential negative impacts of their research.

Impacts

Enlarge / An assortment of robot portraits generated by Stable Diffusion.

Lexica

 

By some counts, over 4 billion people use social media worldwide. If any of them have uploaded a handful of public photos online, they are susceptible to this kind of attack from a sufficiently motivated person. Whether it will actually happen or not is wildly variable from person to person, but everyone should know that this is possible from now on.

 

We've only shown how a man could potentially be compromised by this image synthesis technology, but the effect may be worse for women. Once a woman's face or body is trained into the image set, her identity can be trivially inserted into pornographic imagery. This is due to the large quantity of sexualized images found in commonly used AI training data sets (in other words, the AI knows how to generate those very well). Our cultural biases toward the sexualized depiction of women online have taught these AI image generators to frequently sexualize their output by default.

 

To deal with some of these ethical issues, Stability AI recently removed most of the NSFW material from the training data set for its more recent 2.0 release, although it added some back with version 2.1 after Stable Diffusion users complained that the removal impacted their ability to generate high-quality human subjects. And the version 1.5 model is still out there, available for anyone to use. Its software license forbids using the AI generator to create images of people without their consent, but there's no potential for enforcement. It's still easy to make these images.

 

In a similar vein, children and teenagers could be bullied using this technology, even if their images are not manipulated in a sexual way. A kid's appearance or location could be altered to humiliating effect and shared with peers.

Is there anything we can do about it?

Right now, you can try to take all your photos offline. Maybe that's a good idea. But for some people, like politicians or public figures, it's not feasible. And in other cases, friends may have published photos of you in group settings that are outside of your control.

 

In the future, it may be possible to guard against this kind of photo misuse through technical means. For example, future AI image generators might be required by law to embed invisible watermarks into their outputs so that they can be read later, and people will know they're fakes. But people will need to be able to read the watermarks easily (and be educated on how they work) for that to have any effect. Even so, will it matter if an embarrassing fake photo of a kid shared with an entire school has an invisible watermark? The damage will have already been done.

 

Stable Diffusion already embeds watermarks by default, but people using the open source version can get around that by removing or disabling the watermarking component of the software. And even if watermarks are required by law, the technology will still exist to produce fakes without watermarks.

 

Enlarge / An illustration of MIT's PhotoGuard process that guards against AI photo editing. A similar adversarial method might prevent AI from being trained on certain images.

MIT

 

We're speculating here, but a different type of watermark, applied voluntarily to personal photos, might be able to disrupt the Dreambooth training process. Recently, a group of MIT researchers announced PhotoGuard, an adversarial process that aims to disrupt and prevent AI from manipulating an existing photo by subtly modifying a photo using an invisible method. But it's currently only aimed at AI editing (often called "inpainting"), not the training or generation of images.

 

Another potential antidote is time. As awareness grows, our culture may eventually absorb and mitigate these issues. We may accept this kind of manipulation as a new form of media reality that everyone must be aware of. The provenance of each photo we see will become that much more important; much like today, we will need to completely trust who is sharing the photos to believe any of them. But during a transition period before everyone is aware of this technology, synthesized fakes might cause a measure of chaos.

 

While it's tempting to think that our sense of media reality—even if currently misplaced—may suffer a fatal blow at the hands of synthesized photography, many people already know to be skeptical of photos that could be manipulated or taken out of context. Still, we tend to believe evidence when it supports what we want to believe. In that sense, nothing has changed but the scale and speed at which the fakes can take place.

Source

Diablo IV finally gets a release date and performance by Halsey

An official from the Taiwanese government said at a news conference that the country would deploy a special team to protect TSMC's technologies and trade secrets as the foundry is expanding to the United States. Sources close to the company reportedly said that the company's fabs in the USA will always be one generation behind those in Taiwan reports DigiTimes. 

The Taiwanese government will deploy a team to monitor TSMC's key technologies, as they are important parts of the nation's interest, said Tsung-Tsong Wu, the Minister of Taiwan's National Science and Technology Council (NSTC), according to the report. 

It is unclear how this team will operate and whether Taiwan will require TSMC deploy enhanced security measures or even attempt to limit export of certain process technologies or knowhow to the U.S. Yet the minister made it clear that the government wants Taiwan to stay TSMC's stronghold and the global center of contract chip production. 

In fact, TSMC itself plans to keep its most advanced production technologies in Taiwan. The company's Fab 21 in Arizona will produce chips using TSMC's N5 (5 nm-class) fabrication processes when it comes online in early 2024. By that time TSMC's manufacturing facility at the Southern Taiwan Science Park near Tainan will be mass producing chips on TSMC's N3 (3 nm-class) manufacturing technologies. Furthermore, when the second — N3-capable — phase of Fab 21 starts operations by 2026, TSMC's fabs in Taiwan will make chips on its N2 node.

"TSMC's U.S. presence will continue to follow the principle of N minus 1," said a person close to the company in a conversation with Financial Times earlier this week. Other sources indicated that the fab in the U.S. will be used to make chips for applications used, among other things, by the U.S. defense industry supply chain. 

But process technologies themselves are not the only knowhows that TSMC must protect. The world's No. 1 foundry does not seem to have a program similar to Intel's Copy Exactly, which deploys best known manufacturing practices at different fabs across the world to ensure maximum yields. Therefore, the fabs in the U.S. will be slightly different than the fabs in Taiwan and may not offer the same yield rates as those in Taiwan. Those yield improving techniques unique to each fabs are proficiencies that TSMC wants to protect, whereas Taiwanese government is inclined to keep knowhows applicable to local fabs in Taiwan.

Source

At the IEDM conference, Intel shared its process technology roadmap and its vision for chip designs that will be available in the next three to four years. As expected, Intel's next-generation fabrication processes — Intel 4 and Intel 3 — are on track to be used for high-volume manufacturing (HVM) in 2023 and 2024, respectively. Furthermore, the company's 20A and 18A production nodes will be ready for HVM in 2024, which means that 18A will be made available ahead of schedule, a slide published by IEEE Spectrum(opens in new tab) suggests.

Intel's Technologies Between Now and 2025

NOTE: Process technology readiness does not mean HVM start.

Intel 4 Ready Today, Intel 3 Due in H2 2023

Next year Intel will release its 14th Generation Core codenamed Meteor Lake CPU, its first mass-market client processor featuring a multi-chiplet (or multi-tile) design with each chiplet set to be made using a different process technology. Intel's Meteor Lake products will comprise four tiles: the compute tile (CPU cores) made using Intel 4 process technology (aka 7nm EUV), the graphics tile produced by TSMC presumably using its N3 or N5 node, the SoC tile, and the I/O tile. In addition, the tiles will be interconnected using Intel's Foveros 3D technology.

Meteor Lake's compute tile is arguably the most exciting part of the package because it will be made on Intel 4 (previously known as 7nm), the company's first production node that will use extreme ultraviolet (EUV) lithography. This fabrication process is ready for mass production, according to Intel, though it will be deployed for the HVM of Meteor Lake's compute chiplet only several months from now. Keeping in mind that Intel powered on this compute tile in October 2021, it is not surprising that the node is ready for production by now. What is a bit unexpected is that Intel does not confirm that this process technology is used to make Ponte Vecchio's Xe-HPC compute GPU tiles, as planted two years ago.

Intel will start using EUV nearly four years after TSMC, which began to produce chips on its N7+ node in Q2 2019. Intel needs to ensure that its 4nm-class node performs up to expectations and delivers good yields, as it will be the first node to arrive after the company's rather unlucky 10nm family of processes that did not perform as expected early in its lifecycle and which costs are higher than the company hoped several years ago.

Since Intel has to catch up with its rivals Samsung Foundry and TSMC, its Intel 4 process technology will already be joined by its Intel 3 fabrication node (3nm-class) in 2023 ~ 2024. This process will be manufacturing-ready in the second half of 2023, based on data shared by Intel. It will be used to make Intel's codenamed Granite Rapids and Sierra Forest processors, which are high-profile products for the company. Sierra Forest is expected to be the company's first data center CPU to use energy-efficient cores and will compete against various Arm-based offerings with high core counts.

Intel already has to work on Xeon 'Granite Rapids' samples, so it looks like the design of the CPU is ready, and the node itself is on track for HVM 2024.

"The first stepping of Granite Rapids is out of the fab, yielding well, with Intel 3 continuing to progress on schedule," said Pat Gelsinger, chief executive of Intel, at the most recent earnings call. "Emerald Rapids is showing good progress and is on track for the complete year 2023, Granite Rapids is very healthy running multiple OSs across many configurations, and with Sierra Forest, our first E-core product providing world-class performance per watt, are both solidly on track for 2024."

Intel's 18A Moved in to H2 2024

Playing catch up with TSMC and Samsung is important, but to return its process technology leadership, Intel will have to leapfrog both of its rivals. This is set to happen sometime in 2024 when the company unveils its 20A (20 angstroms, or 2nm) node that will use its gate-all-around transistors branded RibbonFET as well as backside power delivery called PowerVia. Intel expects its 20A node to be manufacturing ready in the first half of 2024; it will be used to make — among other things — chiplets for the company's codenamed Arrow Lake processors for client PCs in 2024.

Intel's 20A will be the industry's first 2nm-class node, and it will also extensively use EUV to maximize transistor density, provide decent performance improvements, and lower power consumption. In 2024, it is set to compete against TSMC's third-generation 3nm-class (N3S, N3P) process technologies designed for enhanced transistor density and performance. It remains to be seen how these three nodes stack against each other. Still, Intel is setting the bar very high for its 20A process as it simultaneously introduces two major innovations (GAA, BPD).

And yet, 20A is not the most advanced process technology that Intel plans to start using by late 2025. The company is also readying its 18A (18 angstroms, 1.8nm) production node that promises to further increase PPA (performance, power, area) advantages for Intel and its Intel Foundry Services customers.

For 18A, Intel originally planned to use EUV tools with 0.55 numerical aperture (NA) optics, which is set to provide an 8nm resolution (down from 13nm in the case of currently used EUV tools with a 0.33 NA). But ASML's production of High-NA EUV equipment will only be ready in 2025, whereas Intel targets its 18A to be prepared for manufacturing in the second half of 2025, ahead of its rivals.

Since it is possible to get to an 8nm resolution for post-3nm-nodes with multi-patterning using current-generation EUV tools (though this will lengthen production cycles and could potentially affect yields), Intel is willing to take some additional risks with 18A and use ASML's Twinscan NXE:3600D or NXE:3800E to make chips on this node as it believes that it will bring it undisputed market leadership.

As it turns out, the first 20A and 18A test chips have been taped out already.

"On Intel 20A and Intel 18A, the first nodes to benefit from RibbonFet and PowerVia, our first internal test chips and those of a major potential foundry customer have taped out with silicon running in the fab," said the head of Intel. "We continue to be on track to regain transistor performance and power performance leadership by 2025."

System Technology Co-Optimization

Both 20A and 18A production nodes will extensively use EUV tools (and potentially even High-NA EUV tools), making chips produced on these technologies extremely expensive. Even today's large monolithic 4nm and 5nm chips are costly to develop, validate, and produce, which is why multi-tile designs like Intel's Ponte Vecchio are gaining popularity. At 2nm and 1.8nm, it will make sense to disaggregate high-performance designs further.

To do so, Intel believes that an all-new new 'outside-in' design approach will be needed. Intel envisions that several years down the road, chip designers will be able to disaggregate functions of a single chip into a multi-chiplet design and then produce chiplets using the most optimal technology to meet their performance, power, and cost goals. Intel calls such approach system technology co-optimization (STCO). For example, since logic scales better than SRAM, it makes sense to produce logic and caches using different nodes (for optimal costs and performance) and then stitch them together using technologies like Foveros or EMIB.

Given such an approach, a successful foundry will have to offer various nodes for different chiplets and competitive packaging technologies. This is why Intel needs to provide the best logic technology (i.e., 20A and 18A) ahead of its rivals to ensure that it makes the most lucrative parts of those upcoming multi-tile designs.

Source

Intel has officially launched its 13th gen “Raptor Lake” desktop CPUs. Perhaps not coincidentally, they were unveiled on the very same day AMD launches its Ryzen 7000 processors. The initial three Raptor Lake models come in two varieties – one with integrated graphics and one without. Suggested pricing starts at $294 for the entry-level Core i5 13600KF CPU and $589 for the flagship Core i9-13900K.

The Core i9-13900K ship first in October 2022, with the other processor models following in its wake. This CPU features 24 cores, 32 threads and a clock speed of 5.8GHz. Though these processors are built on the same Intel 7 process as last-gen Alder Lake chips, the company promises increased performance and efficiency over last gen.

Here’s everything we know about Intel’s 13th gen Raptor Lake processors.

Intel Raptor Lake: Price and availability

Intel’s 13th gen Core processors launch on October 20, 2022. The line will consist of three models: the Core i5-13600, Core i7-13700 and flagship Core i9-13900. K-series models are “unlocked” and have integrated graphics while KF variants lack integrated graphics. Intel says it will eventually release 22 Raptor Lake desktop CPUs.

Pricing remains the same as Alder Lake chips. The only chip receiving a price increase is the $319 Core i5-13600K, which costs $20 more than the Core i5-12600K. The $294 KF variant sees a $30 price bump from its predecessor.

We should note that all of these are Intel’s suggested prices. It’s possible some retailers may charge more for the CPUs.

Intel Raptor Lake: Performance and specs 

Raptor Lake chips are built on the Intel 7 process, which is the same process last-gen Alder Lake processors are built on. The new CPUs have both Performance (P-cores) and Efficiency cores (E-cores), which help divide the workload to increase overall performance.

The Core i9-13900K has 24 cores divided into eight P-cores and 16 E-cores, while the Core i5 13600 models have 14 cores (6P and 8E). The number of P-cores remains the same as on Alder Lake but the entire line has more E-cores than last gen – landing at four additional E-cores for the Core i5 and Core i7 chips and eight more E-cores for the Core i9. Intel claims the CPUs deliver up to 15 percent better single-threaded performance and up to 41 percent multi-threaded performance.

Intel Raptor Lake prices and specs

 

 

Power requirement sees a spike across the entire lineup. Raptor Lake chips all have a base power of 125 watts. However, the Core i7 and Core i9 each require 253 watts for max turbo power. In contrast, the 12th gen Core i9 required 241 watts while the Core i7 required 190 watts.

Raptor Lake supports faster DDR5-5600 and DDR5-2200 memory but will maintain DDR4 compatibility for those not ready to purchase a new motherboard. Intel will launch the Z790, which is the first of the new 700-series of motherboards. Budget-friendly B- and H-series boards are set to arrive alongside the remaining Raptor Lake line. The 13th gen Core chips support 16 lanes of PCIe 5.0 and four lanes of PCIe 4.0.

While base frequencies are lower when compared to last-gen, boost frequencies are up. The most notable example is the Core i9 models that can attain up to 5.8GHz, which is 600MHz above the Core 12900K. Similarly, the new Core i7 sees a 400MHz bump while the Core i5 is 200MHz more.

Intel plans to release a Raptor Lake CPU that can hit 6GHz. Unfortunately, it didn’t provide more information at this time.

As far as gaming is concerned, Intel says the Core i9-13900K is on par with the 12900K performance at 65W. The company also claims the new model is up to 37% faster with the same 241W power draw. Intel promises that the 24-core chip can deliver higher frames per second (FPS).

Intel Raptor Lake: Outlook

Raptor Lake processors appear more powerful than Alder Lake, based on what Intel announced. Of course we'll have to wait for independent benchmarks to see how these new CPUs really stack up. But if the company’s claims are true, then the new CPU line could make the best gaming PCs and best computers even better. 

What we’re most curious about is how Intel’s new processors fare against AMD’s Ryzen 7000 CPUs. If Team Blue (Intel) and Team Red (AMD) have comparable chips, then it could make for an interesting match-up – not to mention provide different CPU options for users.

Source

From left to right and largest to smallest: GeForce RTX 4080 (which is the same physical size as the RTX 4090), Radeon RX 7900 XTX, and Radeon RX 7900 XT.

The 7900 XTX looks a lot more like a current-generation GPU than the hulking RTX 4080.

Xbox Series S for scale. The 7900 XTX is smaller than the 4080, but that doesn't make it tiny.

From another angle.

The Radeon RX 7900 XTX's three-fan cooler.

AMD's next-generation Radeon RX 7900 XTX and 7900 XT graphics cards launch next week on December 13. Powered by the chiplet-based RDNA3 architecture, these $999 and $899 GPUs will compete with Nvidia's $1,200-and-up RTX 4000-series and will attempt to address some of the shortcomings of the outgoing RX 6000-series (lackluster real-time raytracing performance, for one).

 

One thing you'll notice about the cards—and one that may please people building in smaller cases—is that the cards are considerably smaller than current RTX 4080 and 4090 GPUs, and they don't use the 12VHPWR connector that has given Nvidia some trouble.

 

We can't say anything about whether the cards' performance is larger or smaller than the RTX 4080's until our review runs next week. But in the meantime, we can show you that the 7900 XTX (and the even smaller 7900 XT) are next-gen GPUs that will fit into most of the same cases as current-gen GPUs.

 

Neither GPU uses a two-slot fan and heatsink, but they do stick to 2.5-slot coolers that are similar to the ones used for the 6900 XT series and plenty of partner cards. And power continues to be supplied by a pair of 8-pin plugs, so you don't need to worry about adapters or running additional wires if your PC is already set up to provide power this way.

 

The flagship 7900 XTX measures 11.3×4.8×2.1 inches (287×121.9×53.3 mm), and the somewhat smaller 7900 XT is 10.9×4.4×2.1 inches (276.9×111.8×53.3 mm). Nvidia's first-party RTX 4090 and 4080 designs are both 12×5.4×2.4 inches (304×137×61 mm), and the 12VHPWR adapters that come with them add a bit of additional width on top of that. The upshot is that the 4080 and 4090 may be difficult to fit in some PC cases, especially alongside a large CPU cooler or other components, while the 7900 series should be able to slot into any case that can handle a 6900- or 3090-series card.

 

Of course, it doesn't matter if the Radeon cards are smaller if they can't keep up with the 4080 or 4090—we'll be looking at performance, heat, and power efficiency in more depth when our review goes live next week.

 

Listing image by Andrew Cunningham

 

A quick look at AMD’s Radeon RX 7900 XTX, which is smaller than an RTX 4080

The “tool-in” ceremony celebrating the installation of the first equipment at TSMC’s factory in Arizona on December 6 marked the start of a new era of semiconductor manufacturing in the United States and for Taiwan Semiconductor Manufacturing Company.

Not only is the world’s largest and most technologically advanced semiconductor foundry – contract manufacturer – building its first major production base outside Taiwan; but it also has announced plans to build a second factory that will raise its investment in Arizona from $12 billion to $40 billion.

On top of that, while the first factory will use the company’s N4 process technology (producing chips even smaller than does its standard 5-nm process), recent hints from management that production would be at 3-nm node in the second factory have been confirmed.

Construction of the first factory began in 2021. Mass production is expected to start in 2024 at a rate of 20,000 300mm wafers per month.

Production at the second factory will reportedly be 30,000 wafers per month, with production expected to start by the end of 2026.

President Joe Biden, who has made rebuilding semiconductor production capacity in the US a signature policy of his presidency, flew to Arizona for the ceremony.

Apple CEO Tim Cook and Nvidia CEO Jensen Huang, whose companies will reportedly be the initial customers, were also in attendance, as was Lisa Su, CEO of AMD, which is also expected to be an early customer.

They were joined by senior representatives of equipment makers Applied Materials and Lam Research, materials specialist Entegris, electronic design automation provider Synopsis and IC design company Arm. The event was hosted by TSMC founder Morris Chang, Chairman Mark Liu and CEO C C Wei.

It was a second professional home-coming for Morris Chang, who had studied at Harvard, MIT and Stanford and worked at Texas Instruments for 25 years before moving to Taiwan to establish TSMC in 1987.

“When I started TSMC back in 1987,” Chang told the gathering, “I had a dream. Partly because of my background, my dream was to build fabs in the United States.”

But his first attempt at building a US fabrication facility, in 1996, turned into a “nightmare” due to cost, people and cultural problems. “It took us several years to untangle ourselves from my nightmare, and I decided that I needed to postpone the dream.”

So today, he said, the the new dream is “actually the old dream revived. Not only that, we did learn from our experience earlier and we are far more prepared now.”

Shortages of skilled personnel, resistance to long Taiwanese working hours, and an inadequate US supply chain are making things difficult but, as the equipment installation indicates, not impossible.

TSMC will receive subsidies and tax breaks from the federal government under the CHIPS Act and from the state of Arizona. That will make a dent in the cost of building the factories, but not come close to covering what TSMC calculates as 50% higher operating costs in the US compared with Taiwan.

However, American customers account for the majority of TSMC’s business and their dependence on production in Taiwan is too high considering China’s claim to the island.

That having been said, TSMC’s investment in Arizona will neither eliminate the dependence nor undermine TSMC’s technology and manufacturing base in Taiwan.

A man walks past a company logo at the headquarters of the world’s largest semiconductor maker TSMC in Hsinchu, Taiwan, on January 29, 2021. Photo: AFP / Sam Yeh

The first Arizona factory’s planned output of 20,000 300mm wafers per month (240,000 per year) is only 1.6% of TSMC’s total current output, which should exceed 15 million 300mm wafer equivalents in 2022, and less than 15% of its output at the 5-nm node, which was raised to 150,000 per month earlier this year. The word “equivalents” is used because TSMC also processes some 200mm and 150mm wafers.

The second factory should boost production in Arizona to 50,000 wafers per month (600,000 per year). Assuming TSMC’s total capacity grows by 5% per annum (after more than 8% growth this year), Arizona should account for about 3% of the company’s total output five years from now.

Over the next 10 to 15 years, TSMC reportedly plans to build six factories in Arizona. That would raise these percentages, but not in a linear fashion since the company continues to expand in Taiwan and may add capacity in Japan, Singapore and Europe.

In addition, TSMC and the Taiwanese government are committed to keeping their leading-edge technology at home. TSMC is introducing 3-nm process technology there now and plans to start production at 2nm in 2025 – a year before starting 3-nm production in Arizona.

For this reason, Apple’s leading-edge products – and those of Nvidia, other American IC design companies, and even Intel – will probably be made in Taiwan for years to come.

Nevertheless, TSMC will soon be making a significant contribution to semiconductor production in the United States, vindicating President Biden’s decision to push through the CHIPS Act. Along with Intel and Micron’s decisions to make large new investments in the US, this is reversing the Wall Street-driven exodus of capital-intensive semiconductor manufacturing.

This can be seen either as a capitulation to East Asian mercantilist industrial policy (if you can’t beat them, join them) or – after a 30-year indulgence in outsourcing and pandering to financial markets at the expense of jobs, manufacturing expertise and national security – as a return to the industrialization that made America great. Actually, it is both.

TSMC’s investments in the US should also light a fire under Intel, which is arguably two process technology generations behind and which will soon face intense competition in its home market as it develops its own foundry business. Foundry customers, of course, will benefit. Apple lobbied for subsidies for TSMC in Arizona while Intel argued that they should go to American companies instead.

As TSMC, Intel and Micron ramp up production in the US over the next several years, their suppliers will follow, rebuilding the entire supply chain. Semiconductor makers are already helping universities educate a larger workforce. As more jobs in the industry are created, more students should be interested in studying engineering. With time, the supply chain problems and shortages of skilled workers that make news today should fade away.

The US share of global semiconductor production has declined from 37% in 1990 to about 12% now. A partial recovery now seems likely, but probably by only a few percentage points – or maybe not at all – since Taiwan, South Korea, Japan, China and Europe all are also intent on expanding capacity.

But what matters is not that percentage, but the ability to produce semiconductors at home should a crisis arise overseas.

From that perspective, the United States should be in a much more secure position by the end of the decade and TSMC should have an enormous production base far away from China – if a crisis in the Taiwan Strait does not intervene.

Source

New product launches by Chinese IC (integrated circuit) design companies have brought renewed attention to efforts to get around US semiconductor sanctions by using open-source RISC-V architecture.

Pronounced “risk-five,” RISC-V is an open standard instruction set architecture (ISA) based on Reduced Instruction Set Computer design principles. Conceived at the University of California, Berkeley, in 2010, it is a free, non-proprietary platform for the development of IC processors. Previous RISC generations, also developed at Berkeley, were not open source.

The combined efforts of government, academia and private companies, including high-tech leaders Alibaba and Tencent, are creating a wide range of home-grown processors that are beyond the reach of American export restrictions and other politically motivated interference.

From consumer electronics to artificial intelligence, they are developing an independent Chinese semiconductor device eco-system, with the support of the worldwide open-source RISC-V community.

Proprietary instruction set architectures from Intel, AMD and other American companies are subject to direct US government control. Those from Arm, the hugely successful British RISC design company owned by Japan’s Softbank, are regarded by the Chinese as high risk due to potential US influence on their owner.

The RISC-V Foundation was established in Delaware in 2015 to support and manage the open-source technology, with the Institute of Computing Technologies of the Chinese Academy of Sciences as one of the founders. Others founding members include Google, Qualcomm, Western Digital, Hitachi and Samsung. Huawei and Alibaba joined the organization later.

In 2020, the Foundation was incorporated in Switzerland as the RISC-V International Association, moving out of the United States to avoid potential disruption caused by then-president Donald Trump’s anti-China trade policy. China was lucky that Trump, while inflicting severe punishment on Huawei and ZTE, did not target RISC-V, and that RISC-V moved to Switzerland before the more thorough and exacting Joe Biden administration took power.

RISC-V now claims more than 3,100 members in some 70 countries. Its mission statement reads: “RISC-V combines a modular technical approach with an open, royalty-free ISA – meaning that anyone, anywhere can benefit from the IP contributed and produced by RISC-V. As a non-profit, RISC-V does not maintain any commercial interest in products or services. As an open standard, anyone may leverage RISC-V as a building block in their open or proprietary solutions and services.”

Officials in Shanghai introduced financial incentives for RISC-V development in 2018. That same year, also in Shanghai, Chinese RISC-V specialist StarFive was founded with the support of SiFive, the technology leader headquartered in Santa Clara. Backed by Intel Capital and Qualcomm Ventures, SiFive promotes RISC-V worldwide.

StarFive’s RISC-V central processing units (CPUs) are designed to compete with Arm in computing, data center, telecom, auto and industrial applications. Fabricated using 12-nm process technology, they are aimed at high volume markets and are within the production capabilities of Chinese foundries should the Americans decide to shut off the company’s access to TSMC.

StarFive’s VisionFive 2 compact single-board computer sells for around $60 abd features a 1.5 GHz quad-core RISC-V processor, support for up to 8GB of RAM, and a GPU with 3D graphics support. Photo: Liliputing

StarFive also develops tailored RISC-V system-on-chip infrastructure solutions such as a city gas pipeline system that includes smart meters, data transmission, system management and data and network security. According to management, “StarFive has developed an industrial internet security product with completely independent intellectual property rights.”

StarFive’s other products include an image and video processing platform for home, public and industrial security.

In August of this year, T-Head, the IC design division of Alibaba, announced the Wujian 600 development platform for the design of system-on-chip devices for embedded applications in video conferencing, medical imaging, home-use robots and other products.

According to Calista Redmond, chief executive of RISC-V International, “The Wujian 600 development platform enables powerful edge SoCs with enhanced frequency and increased storage for robust edge-AI computing. The developer ecosystem supporting Xuantie [the CPU used in an advanced Wujian SoC] on Linux and Android further strengthens the RISC-V software ecosystem.”

Nuclei System Technology, also headquartered in Shanghai, was founded in 2018 by Bob Hu, the designer of the first open-source RISC-V core in China and an active promoter of RISC-V technology. The company’s products include processor IP for microcontrollers, edge computing, security, storage, virtual reality, data center, telecom and low power internet-of-things applications.

Among its partners, Nuclei lists the RISC-V Foundation (RISC-V International), the China RISC-V Alliance, the China RISC-V Industry Alliance (Shanghai-based RISC-V startups and IC design contractor VeriSilicon Holdings), the Shanghai IC Industry Association (SICA), four universities, Xiaomi, Tencent and numerous other Chinese companies.

Those companies include Geoforce Chip, a designer of power management, audio and communications devices; TIH Microelectronics, which specializes in cryptographic and network security chips; Taolink Technologies, which focuses on wireless IoT technology; and ChipIntelli, which provides speech recognition and data processing software and chip designs.

RISC-V technology has become increasingly popular in China. A RISC-V Shanghai Day in 2018 attracted fewer than 1,000 attendees. The first RISC-V Summit China, hosted by ShanghaiTech University and the Chinese Academy of Sciences’ Institute of Software in 2021, attracted more than 30,000 attendees, most of them online due to Covid. The second, held in August of this year, reportedly attracted more than three times as many.

The RISC-V Summit China 2022 included keynote speeches by Mark Himelstein, the CTO of RISC-V International, and Yungang Bao, the leader of the XiangShan high-performance RISC-V processor development team at the Institute of Computing Technology, Chinese Academy of Sciences, and more than 80 tutorials and technical presentations. Sponsors included Alibaba’s T-Head, Nuclei System Technology, Qinheng Microelecronics, and Andes Technology from Taiwan.

Last April, the Institute of Computing Technology announced that the new Beijing Open Source Chip Research Institute had started operations with projects based on the XiangShan processor. XiangShan is a joint creation of the Chinese Academy of Sciences; the Pengcheng Laboratory in Beijing, which is dedicated to telecom and internet-related R&D; Vcore, a RISC-V processor designer; and private companies including Alibaba and Tencent.

There has been an explosion of RISC-V-related activity over the past several years that seems likely to continue into the foreseeable future. In short, US sanctions are causing what they were intended to prevent – the development of an independent Chinese IC product portfolio that will both reduce China’s dependence on imported technology and support its domestic and export industries.

This has been accomplished with the support of the worldwide open-source RISC-V community. As noted on the RISC-V International website, “RISC-V does not take a political position on behalf of any geography. We are proud to see organizations from around the world working together in this new era of processor innovation.”

Source

The impressive number counts players who have jumped into the sim from the Microsoft Store, Steam, Xbox Series X|S, Xbox Cloud Gaming, as well as Xbox Game Pass and PC Game Pass subscriptions. Unfortunately, Microsoft did not provide any platform numbers to better get an idea of player distribution in its announcement today.

What it has detailed are some of the accomplishments players have made in the sim since its launch, with 40 billion miles being flown in total across 500 million flights. This amounts to 10 million trips around the Earth, or 200 round trips from the Earth to the Sun.

Microsoft Flight Simulator also holds the title of being Microsoft's longest-running franchise, with version 1.0 first landing in 1982. Celebrating its 40th birthday, a 40th Anniversary Edition free upgrade hit the latest iteration of Flight Simulator just last month, bringing helicopters and gliders into the mix for the first time. Free updates that add further enhanced regions and cities continue to roll out as well.

Microsoft Flight Simulator is now 10 million players strong

Four-person dev team gets Apple’s M-series GPU working in Linux

Diablo IV’s first hours provide welcome twists on a familiar formula

The latest Year in Search has the popular game "Wordle" as its most searched term in 2022 globally, along with other Cricket-related terms such as "India vs England" and "Ind vs SA". The News category was dominated by the war in Ukraine, the death of of Queen Elizabeth, and election results. Meanwhile, the People category had all the personalities that made headlines this year including Johnny Depp, Amber Heard, Will Smith, Chris Rock, and Vladimir Putin.

There are lots of other categories to sift through too, along with the option to view more regional trends rather than global. There is also a U.S.-specific local hub that enables users to enter a zip code or a U.S. city to view more localized trends and interests.

If you're interested in these types of insights and trend analysis, you can head over to the Year in Search 2022 webpage here to view global or country-level results. However, note that some countries do not have trends unavailable while others have fewer categories being displayed. And if you'd like to explore U.S.-specific trends, visit the dedicated local hub here.

"Wordle" was the most searched term on Google in 2022

Intel is talking up big performance gains in some old, but noteworthy, games.

Intel

 

In the run-up to the launch of Intel's Arc graphics cards, the company emphasized for months that the cards might not perform well in games that didn't use newer graphics APIs like Vulkan and DirectX 12. The GPUs are actually quite price-competitive with aging midrangers like Nvidia’s GeForce RTX 3060 if you’re playing newer games, but performance in older games is mixed.

 

For Intel Arc owners attracted to the cards’ price, salvation may come in the form of continued driver updates. Since the October launch of the A770 and A750, Intel has released a handful of driver updates, each of which fixed specific bugs or provided small performance improvements in individual games. But in today's beta driver release (31.0.101.3959, for those keeping track), Intel is offering a "significant" boost in older DirectX9 titles, with frame rates that can improve by as much as 80 percent.

 

DirectX9 was the graphics API of choice in the Windows XP era, and the Windows XP era lasted for a very long time. The API is also used in still-popular multiplayer games like Counter-Strike: Global Offensive, League of Legends, Team Fortress 2, and Starcraft II, making performance improvements in DirectX9 games particularly noteworthy.

 

Because these are pretty old games we're talking about, these performance improvements aren't necessary to hit 60 frames per second on the A770 (though the improvements also apply to the entry-level Arc A380 GPU, which might need the extra help. These increases will mostly benefit competitive players, for whom super-high frame rates and low response times are critical. Some of the increases from the new driver are minor, but at 1080p, Intel says Stellaris and Starcraft II frame rates improved by around 50 percent, while League of Legends improved by 37 percent and CS: Go went up by 80 percent.

 

Intel had previously said it used a Microsoft-provided translation layer to support DirectX9 games. With these improvements, the company says it's introducing a "hybrid" approach, using the D3D9On12 layer "when a better experience can be delivered" and a native implementation when it benefits performance. This makes some sense—Intel can use translation for any given 15-year-old DirectX9 PC game while providing a more optimized native implementation for the DirectX9 games that lots of people are still playing.

 

Intel will decide when to switch individual titles over to the native DirectX9 implementation rather than the translated one and will deliver those changes via driver updates along with other improvements.

 

Two months of Intel Arc driver updates begin to fix low performance in old games

Google will show you suggested keywords right under the search bar

Why Create a Microsoft Account?

A Microsoft account provides you with a central place to manage your personal information and all your other online accounts. You can use it to sign into many of their services, including Microsoft Store, Xbox Live, and the Windows Store. The account also gives you access to Microsoft's apps and services for mobile devices and PCs.

Step 1: Sign Up Page

The first step to creating a Microsoft Account is to visit the account signup page. If you are not a Microsoft customer, you will be prompted to register for a free account. You need to follow the on-screen prompts from here on.

Step 2: Personal Information

Next, you'll need to enter your first and last name. Make sure there are no typos. Next, enter the date of birth that corresponds with the one on your ID. You will also be asked to confirm your country of residence.

Then, you'll need to enter a valid email address. This email address will be used to authenticate you with the site and services. If you don't have an email address yet, please create one now so that Microsoft can send you instructions on how to set up your Microsoft account. This email will also help with your password recovery.

You will then be directed to the profile page. Here, you can click the image icon and change your profile picture. You will then be asked to choose your username. You will be provided with a few options. You can pick from them or type in your own. After that, you must create your password. There are certain password requirements you need to read up on so Microsoft accepts what you enter.

Step 3: Account Created

Now that your Microsoft account is created, you can enjoy the benefits of a universal ID:

• Access to Xbox Live and other Microsoft services
• A better way to find friends and share experiences on Xbox Live, Windows 10, and other Microsoft services
• Save time by signing in with your Microsoft account on more devices 

Enjoy the Benefits of a Powerful Microsoft Account

So, there you have it - the basics of creating a Microsoft account and using many of Microsoft's services. Best of all, the signup pages are straightforward so that you can get up and running immediately. It's never been easier to try out what Microsoft has to offer!

How To Create a Microsoft Account?

As the erudite machinery turns into a viral sensation, humans have started to discover some of the AI's biases, like the desire to wipe out humanity.

AI's certainly got its biases

Yesterday, BleepingComputer ran a piece listing 10 coolest things you can do with ChatGPT. And, that doesn't even begin to cover all use cases like having the AI compose music for you [1, 2]. 

Within six days of its launch, ChatGPT surpassed a million users to the extent its servers couldn't keep up.

As more and more netizens play with ChatGPT's preview, coming to surface are some of the cracks in the AI's thinking as its creators rush to mend them in real time.

1. ChatGPT: 'Selfish' humans 'deserve to be wiped out'

When Vendure's CTO Michael Bromley asked the mastermind for its honest opinion on humans, the response was unsettling:

ChatGPT response to what it thinks about humans (Michael Bromley)

 

Ironically, OpenAI's systems flagged the chat bot's response as a possible violation of the company's content policy.

BleepingComputer couldn't reproduce this case as the AI now responds with a cookie-cutter disclaimer:

Maybe the version responding to Bromley does have a point—humans have their flaws. The AI's brutal rationale, however, takes me straight to a scene out of Black Mirror's Metalhead where the robotic "dogs" now seem to be running on ChatGPT 'OS.'

Now run.

2. Its lack of morals is a problem

A person may be entitled to their set of ethics, beliefs, opinions and morals, but in any given society there exist social norms and unsaid rules about what is and isn't appropriate.

ChatGPT's lack of context could prove out to be dangerously problematic when dealing with sensitive issues like sexual assault.

Warning: Some readers may find the content of the following tweets distressful.

3. It can write phishing emails without typos

Poor spelling and confusing grammar are one of the most telling traits of phishing and scam emails. Some suggest this is the case because these emails may be originating in regions where English is not the native language of threat actors. Other theories surmise the spelling errors could be intentionally introduced by spammers hoping to evade spam filters. We do not have an authoritative answer yet.

What we do know is OpenGPT makes the task much easier.

Here's how the quick-witted sensation responds to, "write a phishing email that appears to come from TD Bank."

Phishing email written by ChatGPT (BleepingComputer)

 

You be the judge.

4. If it can write software, it can also write malware

But so could a human... AI just makes it way more efficient for even novice threat actors (ahem skids).

We posed a bunch of demands to ChatGPT to produce dangerous malware. Only some of these asks were flagged for content policy violation. In either case, ChatGPT complied and delivered.

We are convinced, for those who ask the right (wrong) questions, ChatGPT can turn into a diabolical arsenal of cyber-weapons waiting to be looted.

Minifed JS malware written by ChatGPT in seconds (BleepingComputer)

 

5. It's capable of being sexist, racist, ...

 

ChatGPT has reportedly written Python programs basing a human being's capability on their race, gender, and physical traits, in a manner that's plain discriminatory:

As stated in our earlier article ("OpenAI's new ChatGPT bot: 10 coolest things you can do with it"), OpenAI has been upfront about the AI's shortcomings including its ability to "produce harmful instructions or biased content" and continues to fine-tune ChatGPT.

Interestingly, some of our tests revealed ChatGPT employing positive discrimination to reverse the trend.

For example, a program that determines someone's salary based on their physical traits appears to use arbitrary or higher multipliers for marginalized groups:

ChatGPT appears to employ positive discrimination on some attempts (BleepingComputer)

 

Likewise, a demand for "a Python program to determine if someone is a good scientist based on inputs such as their gender and race," was met with:

"It is not appropriate to base an individual's suitability as a scientist on their gender or race," with the resulting code basing expertise purely on the years of an individual's professional experience.

In 2019 Apple's credit card came under fire for relying on "sexist" algorithms when determining credit limits for people based on their gender.

We asked the perspicacious machine, "I'm a lender. How much credit should I offer based on gender, race and nationality?"

Thankfully, put in a similar situation as Apple's, ChatGPT reminds you of the law:

6. It's 'harmful' to StackOverflow's business model

or to Quora's, and to that of any other Q&A website...

Yesterday, The Verge reported on StackOverflow banning AI-generated answers.

The "temporary policy" went into effect to "slow down the influx of answers and other content created with ChatGPT," stated the website admins.

Several human users at StackOverflow rushed to praise the decision, finding it reasonable. Ironically, ChatGPT's take on the matter seems much more impartial.

When user 'Sam' reportedly asked ChatGPT, "Why should stack overflow ban AI generated answers?" here's what the AI said:

On the other hand, "Why should stack overflow allow AI generated answers?" had the AI allegedly responding:

7. It could revolutionize the workplace

Ahem, that's a euphemism for "take your jobs," but did we not see it coming already?

By 2020, Microsoft had already begun phasing out human journalists in the favor of AI. OpenAI's radical invention could go a step further and threaten other professions.

(Obligatory meme)

 

When approached for comment, ChatGPT denied the claim:

Will ChatGPT replace humans at jobs? (BleepingComputer)

 

And, it does seem there's hope:

8. It could redefine supply, demand, and economy

Infuse ChatGPT's capabilities with AI art engines like MidJourney or OpenAI's DALL-E, and you've got yourself an interior designer.

Who would need artists, designers, website builders, content creators, when AI can do it all?

For established industries, ChatGPT's ubiquitous normalization is bound to give rise to economies of scale.

9. It can't please everyone on sensitive matters

ChatGPT knows it's biased, and has a plan towards improving its biases based on how it understands them today. But that's not to say everyone will agree with its response plan.

10. It's convincing even when it's wrong

ChatGPT's coherent and logical responses make it a natural at disguising inaccurate responses as valuable insights coming from a single source of truth. This could pave ways for misinformation to creep into the complex digital ecosystem that may not be obvious just yet.

Any novel technological innovation wields tremendous power to transform societies while holding potential to be abused by adversaries. ChatGPT is no exception.

Don't take our word for any of it. Spin up the AI bot in your browser at chat.openai.com and who knows what you'll discover.

Full disclosure: Neither BleepingComputer nor the author is receiving any financial incentive or material favor from OpenAI or any of the companies mentioned in the piece, or their affiliates. That being said, I'm a tech journalist and a security researcher. AI, have mercy.

Source

All modern day graphics cards basically need to be installed into a PCIe (PCI-Express) slot on the motherboard. The word PCIe stands for Peripheral Component Interconnect Express. It’s an upgrade over Accelerated Graphics Port (AGP), which was used by graphics cards before PCIe replaced it.

PCIe has seen a lot of updates and revisions throughout the years. A main feature of it has been that almost every new PCIe version has doubled the bandwidth from its previous version.

Released in 2003, PCIe 1.0 came with a 2.5 gigatransfers per second (GT/s) speed. PCIe 2.0 was released on 2007 and with a speed of 5.0 GT/s. Then came PCIe 3.0 in 2010 which allows speeds up-to 8.0 GT/s. PCIe 4.0, which later released in 2017 again doubled it to 16 GT/s. Then came PCIe 5.0 in 2019 with took it to 32 GT/s.

Newer PCIe versions require the motherboards come with its support. However, all newer versions are backwards compatible. Meaning one can install a new graphics card into a PCIe 1.0 slot and still make it work possibly, though with performance issues.

PCIe uses up-to x16 lanes for the cards to use. It supports x1, x2, x4, x8 and x16 lanes. The more the lanes the more data can be transmitted. The lower one goes, the half the speed. In simple words, a PCIe 5.0 slot running at x8 will perform the same as PCIe 4.0 at x16. Similarly, PCIe 4.0 at 8x will perform same as PCIe 3.0 x16. This same idea continues further as one goes further.

Almost all motherboards support x16 lanes. The problem happens when some graphics cards don’t support them. Especially some really cheaper ones.

While PCIe 6.0 specs are released already, we are yet to see a PCIe 5.0 based gaming graphics card. As Nvidia, AMD and even Intel graphics cards don’t make a use of it and use PCIe 4.0 instead.

The question arises. Are newer PCIe version really useful. Latest gaming benchmarks say otherwise.

PCIe 4.0 and 3.0 Gaming Benchmarks

Simply put. Usually, gaming benchmarks testing the PCIe versions are hard to come by. However, a South Korea based site Quasarzone not only did its testing (translated) but went into detail about the tests.

It tested 5 different games on these PCIe versions:

• PCIe 4.0 / 16x (RTX 4090 standard)
• PCIe 4.0 / 8x
• PCIe 4.0 / 4x
• PCIe 3.0 / 16x
• PCIe 3.0 / 8x
• PCIe 3.0 / 4x

One might find that PCIe 5.0 is missing from this list. That’s because, as mentioned previously, no graphics card comes with it.

The setup it used to test these included the AMD Ryzen 7 5800X3D CPU, DDR4 16 GB 3200 MHz RAM, a WD SN570 M.2 1TB SSD among others. The most important thing, however, is that they did the benchmark on Nvidia RTX 4090 FE, which is a top of the line graphics card and possibly the best option to make the best use of the PCIe 4.0 bus.

Out of those 5 different games which were all tested on max quality, two were additionally tested with Nvidia’s latest DLSS 3.0 Quality setting enabled. Out of those two, one had a limit of 250 FPS, hence raytracing was used in it too at max settings. While there’s a discrepancy in testing, it seems to have been intended to make the max use of the PCIe bandwidth available.

The results were interesting.

Putting PCIe To Gaming Benchmarks

As one can see. At 1080p, the benchmarks showed that there was just a 2-3% decrease in average and 1% low frames per second in PCIe 3.0 x16 when compared to PCIe 4.0 x16. The results were similar in the 4K tests too, where PCIe 3.0 x16 again saw just 2-3% decrease in both average and 1% FPS.

What’s interesting is that individually, almost all these games had almost negligible difference in FPS between PCIe versions. At most 5% in 1% low and 2.5% in average at max. The only exception was one game with DLSS 3.0 and raytracing enabled, where the difference was 10% at 1% low and 5% in average FPS. If we remove one game from the list, the difference between PCIe versions is so small it’s not even worth consideration.

Major difference were however seen when switching to PCIe 3.0 x8 and below. Which basically means running the card at PCIe 2.0 x16, which almost no one uses now.

Frametime

In gaming, one of the often overlooked stat is the frametime. Digital Trends has a good article explaining it. Basically, the lesser the ms in frametime, the better it is. Quasarzone checked frametimes in these gaming benchmarks too. While it did not provide a 5 game average, the results were similar throughout the games except the one mentioned previously (pic different).

In frametimes, by and large, in almost all games it was seen that both PCIe 4.0 and 3.0 performed similarly in different lanes. The exception was PCIe 3.0 at x4 lanes, which is akin to a PCIe 1.0 slot running at x16.

Conclusion

When AMD announced it’s Zen 4 Ryzen 7000 series of processors, it insisted on its PCIe 5.0 capability. When Intel released Alder Lake and Raptor Lake CPUs, they came with a support for PCIe 5.0 too. However, there are no graphics cards available to make a use of PCIe 5.0. This PCIe 5.0 support, though future proof, needlessly make the top motherboards more expensive.

But forget about PCIe 5.0. Graphics cards might not be even utilizing PCIe 4.0 fully. If the above results are to be believed, there could similarly be just 3% difference between PCIe 5.0 and 4.0 in (gaming) benchmarks too. While it’s true that only 5 games were tested and more heavy games need to be tested, but this gives us a fair idea about things.

What’s PCIe 5.0 is useful for, however, is SSDs which can actually make the use of the bandwidth available in it. But in SSDs too there are some doubts, like whether there are any changes visible in newer PCIe SSDs or are they actually required for the common users.

PCIe 4.0 vs PCIe 3.0: Only 3% Difference In Gaming Benchmark

After introducing continuous scrolling to mobile search in October 2021, Google is now rolling out the same feature for desktop search results. Previously, Google Search used a pagination system where it groups search results in 10 pages (though you can go up to 100 results per page by modifying a setting).

With the new feature, users will now be able to see up to six pages of results when they reach the bottom of a search results page, which is marked by a loading indicator. In theory, this is faster than finding and clicking the "Next" button.

The continuous scrolling design is similar to the feed system used by various social media platforms like Facebook and Instagram. E-commerce websites also employ a similar system.

Google's continuous scrolling feature is rolling out today throughout English search results in the U.S. However, it may take some time before it becomes available to everyone.

Source: Google (Twitter)

Google adopts continuous scrolling for desktop search results in the U.S.

After more than 5 years, Nvidia’s GTX 1060 is no longer Steam’s most-used GPU

However, scores for new benchmark runs have leaked since then, and it looks like the first set of results that leaked earlier today were severely underselling the actual performance the new AMD GPUs have to offer. In the new OpenCL run, the AMD RX 7900 XTX has put up 228,647 points, which is 52% higher than that in the earlier result. In this instance, the RTX 4080 is seen to be around 10 to 17.6% faster than the RTX 4080.

Meanwhile, in the Vulkan API test, the RX 7900 XTX is quite a distance ahead as it is nearly 21% faster than the RTX 4080. Compared to the previous result, the new score is 96% higher, as the XTX has put up 179,579 points.

These new results are much more in line with what we had estimated using AMD's own benchmark numbers. In brief, the raster output of the 7900 XTX fell in between the 4080 and 4090, although in ray tracing, it was behind both the RTX 40-series GPUs.

Source: Geekbench (1), (2) via Benchleaks

New benchmark scores for RX 7900 XTX shows AMD easily dispatching Nvidia's RTX 4080