Graphics card manufacturer eVGA has made a name for itself manufacturing and selling Nvidia's GeForce GPUs for two decades, including some of the more attractively priced options on the market. But according to the YouTubers at Gamers Nexus, analyst Jon Peddie, and an EVGA forum post, EVGA is officially terminating its relationship with Nvidia and will not be manufacturing cards based on the company's RTX 4000-series GPUs.

EVGA's graphics cards have exclusively used Nvidia GPUs since its founding in 1999, and according to Gamers Nexus, GeForce sales represent 80 percent of EVGA's revenue, making this a momentous and arguably company-endangering change. But EVGA CEO Andrew Han told Gamers Nexus that the decision was about "principle" rather than financials—Han complained about a lack of communication from Nvidia about new products, including information about pricing and availability.

Nvidia's pricing strategy was apparently another sore point for EVGA. Nvidia's first-party Founders Edition cards could often undercut the pricing of cards offered by EVGA and other vendors, forcing them to either lower prices or lose sales as a result.

Nvidia may not be entirely at fault here—the wider dynamics of the GPU market are also tough to navigate. As Peddie points out, even as GPU costs have gone up, profit margins for the board partners that manufacture Nvidia graphics cards have gone down. Modern high-end graphics cards have massively higher power, cooling, and PCI Express signaling requirements than cards from just a few years ago, making them more expensive to design and manufacture, and reporting about the RTX 4000 series indicates that that trend is only going to continue.

And it probably doesn't help that the GPU market has been plummeting back to earth this year, after over a year of limited stock and inflated pricing. Sliding cryptocurrency values and the Ethereum cryptocurrency's move away from GPU mining have both flooded the market with secondhand GPUs, which has in turn affected demand for new GPUs. In Nvidia's last earnings call, CEO Jensen Huang complained of "excess inventory" of RTX 3000-series GPUs that made it miss its quarterly revenue projections by $1.4 billion.

For Nvidia's part, its public stance can be summed up as "so long and good luck."

"We’ve had a great partnership with EVGA over the years and will continue to support them on our current generation of products," Nvidia spokesperson Bryan Del Rizzo told Tom's Hardware. "We wish Andrew [Han] and our friends at EVGA all the best."

Earlier this week, Google made some serious cuts to its startup incubator Area 120, cutting half of its projects, according to TechCrunch. The purpose of Area 120 was to give Google employees somewhere to experiment or chase their passion projects, with the hope that they could stumble upon the next Big Idea like Adsense, Gmail, or Google News.

But as the economy has turned, it seems Google may be losing its stomach for big bets and experimentation, instead trying to focus its efforts on what makes money today (or services that it really just has to run despite the fact that they lose money, like Google Cloud). Instead of challenging employees to spend 20 percent of their time building wacky apps, CEO Sundar Pichai says the company needs to be 20 percent more efficient and productive. In a memo earlier this summer, he urged employees to be “more entrepreneurial,” but it read more like a demand to work harder and find ways to cut costs rather than to bark up trees that may or may not bear future fruit.

That perception has permeated outside the company as well. According to a recent report from The Information, some recruiters looking to hire employees for startups are starting to look away from Google because they have the impression that the tech giant’s employees mostly maintain legacy products instead of building new ones.

One of the places this idea is most evident is the company’s hardware. Earlier this week, we reported Google canceled a Pixelbook project that was “deep into development” to cut costs, effectively leaving it up to other companies to push the Chromebook category forward.

As my colleague Monica Chin pointed out, the original Pixelbook felt more like a halo device meant to inspire other manufacturers and show what was possible with Chromebooks rather than something Google actually expected people to buy. While we don’t know whether the canceled Chromebook would’ve had that same ethos, it feels safe to say that launching any sort of laptop wouldn’t be a guaranteed home run for Google — according to market analysis firm IDC, the company isn’t even one of the top five Chromebook manufacturers in terms of market share or units shipped. Google would have to convince people to choose its laptop over ones from trusted brands like Acer, HP, or Lenovo, something it’s seemingly failed to do at any sort of scale with the 2019 Pixelbook Go (perhaps owing to its rather high $649 starting price).

Other parts of Google’s Pixel lineup also feel less bold than they used to. Remember the Pixel 4, which came with a radar sensor, or the Pixel 2, which introduced squeezable sides to Google’s phones? Those were fun, cool features that gave you a good reason to consider a Pixel. Everything we’ve seen about the Pixel 7 so far makes it look like it’ll be a relatively minor upgrade, complete with a very similar design to its predecessor. We’ll find out if that’s actually the case on October 6th, but I won’t be shocked if there’s nothing shocking about Google’s next phones.

It’d be unfair to say that Google isn’t doing anything new when it comes to Pixels, though. It is, after all, adding both a smartwatch and a tablet to its lineup. However, neither seem to be pushing any boundaries, especially the latter, which has a design that feels years out of date now and will likely feel even more so next year after we get a fresh new crop of iPads. Instead of being halo devices that inspire other manufacturers, it just feels like Google’s playing catch-up to Samsung (and at this point, it’s unclear if these devices will even be competitive).

This phenomenon isn’t entirely new for Google, and it’s not just limited to hardware. Google is infamous for giving up on projects shortly after they launch, and sometimes even before — just in the past year or so, it’s turned around on plans to integrate banking into Google Pay (which has largely been replaced now by a reincarnated Google Wallet), mostly done away with its YouTube Originals program, and moved Stadia away from being a game streaming service to more of a white-label tech that companies can use for demos and cell plan add-ons. But there’s a difference between giving up on select projects and shifting your culture to be more conservative towards experimentation. Declaring for certain which of those Google’s doing would require more data, but moves like transferring assets from a failed experiment to a startup and then investing in the company feel less ambitious than just rebooting the project yourself.

None of this is to say that Google is completely standing still — obviously, its core apps and services are continuing to get new features, redesigns, and tweaks. And it’s not like Google’s stopped throwing everything it can at the messaging platform wall. The company spends almost $10 billion a quarter on research and development — that money is obviously going somewhere. But I’d be hard-pressed to think of any of the company’s recent work that really made me sit up and go, “Wow!” Sure, I enjoy watching the occasional TikTok but on Youtube, and I appreciate that Google’s making Android more customizable and web tracking a little less creepy, but those changes are incremental, not revolutionary.

Perhaps part of the reason it’s been hard to get excited about Google’s efforts is where it’s focusing on innovating. It’s introduced a flurry of changes to Workspace recently, adding “chips” that let you blend your documents, spreadsheets, reminders, and even meetings and emails together. It’s also been paying some attention to Meet, its Zoom competitor. But while these changes have probably made people’s work lives a little easier, adding new features in an office suite doesn’t exactly seem like swinging for the fences to build the future to me.

Even if Google were to completely stop all experimentation, it would likely survive — its services are engrained in the fabric of how almost all of us use the Internet at this point. But if it doesn’t take big bets, it’ll be hard for it to come up with the next Gmail, Google Assistant, or ChromeOS and almost impossible to help invent new categories of tech like self-driving cars or ambient computing. If companies want to attract the sort of people who are going to build the future, they have to be the type of place where people can actually go out on limbs and not be worried about getting in trouble for barking up the wrong tree. It’d be a shame if Google became a company where that wasn’t the case.

Ethereum developers today executed the "Merge," an upgrade that eliminates mining and dramatically reduces the energy consumption of the world's second-biggest cryptocurrency. Today's action "completed Ethereum's transition to proof-of-stake consensus, officially deprecating proof-of-work and reducing energy consumption by ~99.95 percent," the Ethereum.org Merge page said.

The Ethereum blockchain has existed since July 2015, and planning for today's change has been in the works for several years. Because a botched transition could have caused chaos, Ethereum developers over the past year have "repeatedly pushed back the date of 'the Merge' to give themselves more time to prepare," as Ars writer Timothy B. Lee previously wrote in a detailed feature on the transition. The Merge will "put the world's Ethereum miners out of work," as the new system doesn't require the powerful graphics cards previously needed to maintain the blockchain and create new ether, Lee wrote.

The switch was highly anticipated. "When the Merge officially kicked in at 6:43 a.m. UTC, more than 41,000 people were tuned in on YouTube to an 'Ethereum Mainnet Merge Viewing Party,'" CoinDesk wrote. "They watched with bated breath as key metrics trickled in suggesting that Ethereum's core systems had remained intact. After about 15 long minutes, the Merge officially finalized, meaning it could be declared a success."

Before the Merge, Ethereum's annualized power consumption was comparable to the country of Chile's, and its carbon footprint was similar to Hong Kong's, according to Digiconomist's Ethereum Energy Consumption Index.

The ether price was down almost 9 percent today as of this writing, while bitcoin had dropped about 2.4 percent.

No more mining

The official Ethereum website explains that the Merge "was the joining of the original execution layer of Ethereum (the Mainnet that has existed since genesis) with its new proof-of-stake consensus layer, the Beacon Chain. It eliminated the need for energy-intensive mining and instead enabled the network to be secured using staked ETH."

The Beacon Chain was created in December 2020 "as a separate blockchain to Mainnet, running in parallel." After a lot of testing, it was ready to take over.

"The Beacon Chain was not originally processing Mainnet transactions. Instead, it was reaching consensus on its own state by agreeing on active validators and their account balances," the Ethereum.org merge page says. "After extensive testing, it became time for the Beacon Chain to reach consensus on real world data. After The Merge, the Beacon Chain became the consensus engine for all network data, including execution layer transactions and account balances."

Now that the change is complete, "mining is no longer the means of producing valid blocks. Instead, the proof-of-stake validators have adopted this role and are now responsible for processing the validity of all transactions and proposing blocks." The joining of Mainnet with the Beacon Chain "also merged the entire transactional history of Ethereum," so no history was lost in the process.

The change should be seamless for people who hold ether. Funds will still be accessible without any user action. "There is no such thing as 'old ETH'/'new ETH' or 'ETH1'/'ETH2' and wallets work exactly the same after The Merge as they did before—people telling you otherwise are likely scammers," the Ethereum project said.

Less ether will be issued

Another Ethereum.org page explains how the issuance of ether will change post-Merge and why less ether needs to be issued:

Validators on the Beacon Chain are rewarded with ETH for attesting to the state of the chain and proposing blocks. Rewards (or penalties) are calculated and distributed at each epoch (every 6.4 minutes) based on validator performance. The validator rewards are significantly less than the miner rewards issued on proof-of-work (2 ETH every ~13.5 seconds), as operating a validating node is not an economically intense activity and thus does not require or warrant as high a reward.

By contrast, "mining is an economically intensive activity, requiring high levels of ETH issuance to sustain," the page says. Before the Merge, mining rewards totaled about 13,000 ETH per day, and rewards for staking were 1,600 ETH per day.

"After The Merge, only the ~1,600 ETH per day will remain, dropping total new ETH issuance by ~90 percent," the page says. To participate, "validators explicitly stake capital in the form of ETH into a smart contract on Ethereum," according to the proof-of-stake explanation. "This staked ETH then acts as collateral that can be destroyed if the validator behaves dishonestly or lazily."

A validator must deposit 32 ETH into the deposit contract and run software including an execution client, a consensus client, and a validator.

"Whereas under proof-of-work, the timing of blocks is determined by the mining difficulty, in proof-of-stake, the tempo is fixed," the proof-of-stake page says. "Time in proof-of-stake Ethereum is divided into slots (12 seconds) and epochs (32 slots). One validator is randomly selected to be a block proposer in every slot. This validator is responsible for creating a new block and sending it out to other nodes on the network. Also in every slot, a committee of validators is randomly chosen, whose votes are used to determine the validity of the block being proposed."

The research community is excited about the potential of DNA to function as long-term archival storage. That's largely because it's extremely dense, chemically stable for tens of thousands of years, and comes in a format we're unlikely to forget how to read. While there has been some interesting progress, efforts have mostly stayed in the research community because of the high costs and extremely slow read and write speeds. These are problems that need to be solved before DNA-based storage can be practical.

So we were surprised to hear that storage giant Seagate had entered into a collaboration with a DNA-based storage company called Catalog. To find out how close the company's technology is to being useful, we talked to Catalog's CEO, Hyunjun Park. Park indicated that Catalog's approach is counterintuitive on two levels: It doesn't store data the way you'd expect, and it isn't focusing on archival storage at all.

A different sort of storage

DNA is a molecule that can be thought of as a linear array of bases, with each base being one of four distinct chemicals: A, T, C, or G. Typically, each base of the DNA molecule is used to hold two bits of information, with the bit values conveyed by the specific base that is present. So A can encode 00, T can encode 01, C can encode 10, and G can encode 11; with this encoding, the molecule AA would store 0000, while AC would store 0010, and so on. We can synthesize DNA molecules hundreds of bases long with high efficiency, and we can add flanking sequences that provide the equivalent of file system information, telling us which part of a chunk of binary data an individual piece of DNA represents.

The problem with this approach is that the longer the string of bits is that you want to store, the more time and money it takes. Robotic hardware performs the synthesis reactions, and each hardware unit can only synthesize a single DNA molecule at a time. The raw materials the hardware uses to perform that synthesis also add a cost for each stored molecule. While this isn't a concern for small-scale demonstration projects, the costs quickly become prohibitive if you start storing large amounts of data. Citing a DNA synthesis cost of about .03 cents per base, Park said, ".03 cents times two bits per base pair times, say, gigabytes—that's a lot of money. That's millions of dollars."

Park told Ars that Catalog started by rethinking the encoding process to get around this bottleneck. The company's encoding starts with a library of dozens to hundreds of short pieces of DNA called oligos (short for oligonucleotide). Each bit in the data is then assigned a unique combination of oligos—you can think of this as a bit like a silicon processor assigning a bit in memory a unique, 64-bit address. If that bit is a 1, a robot can gather small samples of solutions containing each of the oligos needed to represent it and combine them with an enzyme that can link all of the oligos together.

The enzyme merges the oligos into a single, longer DNA molecule that contains the unique signature of the bit. If, in contrast, the bit is a zero, the corresponding DNA for its address isn't synthesized.

All of the molecules that are produced can then be combined in a single solution (which can be dried out for long-term storage). To read the data, the population of DNA molecules is sequenced, and an algorithm recognizes the unique combination of oligos present in each molecule. The recognized addresses are assigned a 1; the rest, a 0. This restores the data that was encoded to digital form.

This system is far less efficient in data/DNA than storing two bits in every base. But the individual molecules remain small enough that it's still an impressively compact and stable storage medium. And it saves significant time and money due to a fundamental asymmetry: It's far cheaper to synthesize a lot of one specific DNA sequence than it is to synthesize small amounts of lots of different DNA sequences. So by assembling DNA using a small bit of a large volume of pre-made DNA, the cost of synthesis goes down dramatically. Each assembly reaction can also be run in parallel; in contrast, synthesizing individual sequences ties up the machine they're running on until the synthesis is complete.

Backup and cloud storage company Backblaze has published data comparing the long-term reliability of solid-state storage drives and traditional spinning hard drives in its data center. Based on data collected since the company began using SSDs as boot drives in late 2018, Backblaze cloud storage evangelist Andy Klein published a report yesterday showing that the company's SSDs are failing at a much lower rate than its HDDs as the drives age.

Backblaze has published drive failure statistics (and related commentary) for years now; the hard drive-focused reports observe the behavior of tens of thousands of data storage and boot drives across most major manufacturers. The reports are comprehensive enough that we can draw at least some conclusions about which companies make the most (and least) reliable drives.

The sample size for this SSD data is much smaller, both in the number and variety of drives tested—they're mostly 2.5-inch drives from Crucial, Seagate, and Dell, with little representation of Western Digital/SanDisk and no data from Samsung drives at all. This makes the data less useful for comparing relative reliability between companies, but it can still be useful for comparing the overall reliability of hard drives to the reliability of SSDs doing the same work.

Backblaze uses SSDs as boot drives for its servers rather than data storage, and its data compares these drives to HDDs that were also being used as boot drives. The company says these drives handle the storage of logs, temporary files, SMART stats, and other data in addition to booting—they're not writing terabytes of data every day, but they're not just sitting there doing nothing once the server has booted, either.

Over their first four years of service, SSDs fail at a lower rate than HDDs overall, but the curve looks basically the same—few failures in year one, a jump in year two, a small decline in year three, and another increase in year four. But once you hit year five, HDD failure rates begin going upward quickly—jumping from a 1.83 percent failure rate in year four to 3.55 percent in year five. Backblaze's SSDs, on the other hand, continued to fail at roughly the same 1 percent rate as they did the year before.

This data—both the reliability gap between them and the fact that HDDs start to sputter out sooner than SSDs—makes intuitive sense. All else being equal, you'd expect a drive with a bunch of moving parts to have more points of failure than one with no moving parts. But it's still interesting to see that case made with data from thousands of drives over a few years of use.

Klein speculates that the SSDs "could hit the wall" and begin failing at higher rates as their NAND flash chips wear out. If that were the case, you'd see the lower-capacity drives begin to fail at a higher rate than higher-capacity drives since a drive with more NAND has a higher write tolerance. You'd also likely see a lot of these drives start to fail around the same time since they're all doing similar work. Home users who are constantly creating, editing, and moving around large multi-gigabyte files could also see their drives wear out faster than they do in Backblaze's usage scenario.

For anyone who would like to poke at the raw data Backblaze uses to generate its reports, the company makes it available for download here.

Twitter shareholders voted to approve Elon Musk's purchase of the company, weeks ahead of a trial over Musk's attempt to exit the merger deal. Though a specific vote tally wasn't available today, multiple news reports said investors backed the Twitter board's recommendation to approve the $44 billion deal that Musk agreed to in April before changing his mind.

"A majority of Twitter shareholders voted in favor of accepting Musk's $54.20-a-share offer to acquire the social-networking company, according to a preliminary vote count read on Tuesday," Bloomberg wrote.

Today's shareholder vote was the last remaining approval Twitter needed for the Musk deal, but the bigger question is what will happen at the upcoming trial at Delaware Court of Chancery. Twitter sued Musk to force him to complete the deal, and a trial is scheduled to begin on October 17.

Musk owns about 9.2 percent of Twitter stock and wasn't expected to vote "given that he has alleged that Twitter breached the merger agreement," The Wall Street Journal wrote. "The agreement requires Mr. Musk to vote his shares in favor of the deal, though his support isn't crucial if enough other investors back it."

Twitter stock was up 0.7 percent today despite a big drop in the overall market. Twitter was at $41.70 at closing, and shareholders would receive $54.20 per share if Musk has to complete the purchase.

Today's "shareholder meeting lasted 7 minutes, with polls open for about 3 minutes," the Bloomberg article said. "Shareholders could also submit votes for several weeks ahead of the meeting." News reports before today's meeting indicated there were already enough votes to approve the merger.

Judge criticized Musk ahead of trial

Musk has tried to exit the merger by claiming Twitter lied about the number of spam bots on its service. Musk has repeatedly complained about the overall number of bots on Twitter but hasn't disproven Twitter's specific estimate, which is that less than 5 percent of its monetizable daily active users (mDAU) are spam or fake.

Musk lost some key rulings in the pre-trial phase. His attempt to delay the trial until February 2023 was rejected in July. Last week, Judge Kathaleen McCormick rejected Musk's newer motion for a four-week delay, writing in her ruling that "even four weeks' delay would risk further harm to Twitter too great to justify."

McCormick last week also criticized Musk for failing to provide documents sought by Twitter, writing that "Musk's own production of text messages revealed glaring deficiencies." She ordered Musk to produce more documents, noting that Twitter "has born[e] the bulk of the burden of discovery" while "Defendants had less to do but still fell short in their obligations."

A Twitter filing made public yesterday said that gaps in Musk's production of text messages "are notable because they correspond precisely to the period when Musk apparently developed buyer's remorse and set into action his scheme to escape the merger agreement."

More Musk texts made public

It's clear that Musk's document "production was incomplete because other parties have produced messages to and from Musk during this time period that Musk should have produced," Twitter wrote. That includes texts between Musk and Morgan Stanley's head of global technology investment banking, Michael Grimes, which were produced by Morgan Stanley.

Those messages show that on May 8, Musk wrote to Grimes that he was thinking of exiting the merger agreement due to the possibility of "World War III." About a half hour later, according to Twitter's newly public filing, Musk wrote to Grimes that "An extremely fundamental due diligence item is understanding exactly how Twitter confirms that 95% of their daily active users are both real people and not double-counted."

"If that number is more like 50% or lower, which is what I would guess based on my feed, then they have been fundamentally misrepresenting the value of Twitter to advertisers and investors," Musk wrote to Grimes. "To be super clear, this deal moves forward if it passes due diligence, but obviously not if there are massive gaping issues."

Twitter's filing said these messages came weeks after Musk "expressly disclaimed any diligence before signing and agreeing to a no-diligence merger."

Intel is working on making its Thunderbolt technology support a throughput of 80Gbps. That would put it on par with the upcoming USB4 Version 2.0 specification and double the throughput of today's Thunderbolt 4.

Intel showed a video of the "first test system" using a Thunderbolt connection running at 80Gbps today, Tom's Hardware reported from the Intel Development Center in Haifa, Israel. The demo was not live, and the company didn't provide a name or release date for the next generation of Thunderbolt, so we expect it to be a while before consumers start seeing this kind of performance in products.

The Thunderbolt connection reportedly hit 80Gbps by use of two Thunderbolt lanes, with each one running at 40Gbps. With USB-C becoming increasingly popular among gadgets and some countries considering, pushing for, or issuing USB-C mandates for electronics, it's not surprising to hear that the demo ran over the slim connector.

That's all we know about Intel's 80Gbps Thunderbolt, but the USB-IF has said that USB4 Version 2.0 will double the throughput of USB4 by using a new physical layer architecture. USB4 Version 2.0's 80Gbps operation is supposed to work over upcoming 80Gbps active USB-C cables, as well as 40Gbps passive USB-C cables that you can buy today.

USB-C technologies' boost to 80Gbps could be a boon for users with extreme needs, bringing the potential to boost performance with things like external storage, eGPUs, and high-resolution monitors.

We'll stay tuned for any further Thunderbolt developments, and we should learn more about how USB4 Version 2.0 operates at 80Gbps in November, when the USB-IF publishes the spec.

For years, Amazon's Kindle Paperwhite has been the best Kindle for most people to buy, compared to the standard no-adjective Kindle. It was the first to get a front-lit, high-resolution screen; waterproofing; and USB-C... the list goes on. But today, Amazon announced a new $100 Kindle that includes almost all of the best features from the latest Paperwhite for $40 less. The new device is available for preorder now and will be released on October 12.

Compared to the previous-generation Kindle, the new 11th-generation model bumps the pixel density from 167 PPI up to 300 PPI, the same level as Amazon's other Kindles. This makes text appreciably sharper and nicer-looking. Storage is doubled from 8GB to 16GB, not something you'll notice if you primarily read text but more useful if you listen to lots of audiobooks. And the micro USB charging port is gone, replaced by a modern USB-C port that will let you charge the device using the same connector you use for most modern laptops, phones, tablets, and game consoles.

The new Kindle is $10 more expensive than the one it replaces, and the price goes up to $120 if you want the version with no lock screen ads. A Kindle Kids edition with an extended two-year warranty, one of three colorful covers, and a one-year subscription to Amazon's Kids+ service is also $120 (it won't show your kids ads by default, but Amazon says it will display ads if you exit Kindle Kids mode).

The Paperwhite does still have some important advantages over the basic Kindle. The front light for its larger 6.8-inch screen uses 17 LEDs instead of the Kindle's four for smoother and more uniform lighting. The color temperature of its front light can be adjusted if you prefer a warmer look. The Paperwhite Signature Edition includes up to 32GB of storage, an auto-brightness sensor, and wireless charging. And most importantly, all Paperwhite models are IPX8 waterproof, for use in the bathtub, in a swimming pool, or at the beach.

For most people, those improvements are probably still worth the extra $40. But this upgrade to the basic Kindle brings it in line with where the Paperwhite was a couple of years ago, and for people who want to save money (or people who are buying a second Kindle for a kid), it's nice to see it get such a significant upgrade.

Ars Technica may earn compensation for sales from links on this post through affiliate programs.

Earlier this year, news emerged about yet another Atari classic-gaming compilation meant to celebrate the company's 50th anniversary. As we've seen quite a few Atari-branded collections over the years, we wondered exactly what shape Atari 50's selection of "over 100 games" would take ahead of its launch on Steam, Xbox One, PS4, and Nintendo Switch on November 8.

Thanks to a European retailer listing that emerged over the weekend, we now have an apparently final list of the Atari 50 selection of games: 103 in all, as spread across arcade cabinets, six console families, and a selection of "reimagined," newly coded games and ports. The verdict: It's pretty good, yet it's still a glaring reminder that the compilation might better be named "Atari Corp. 50."

Missing vowels and details about Vctr Sctr (for now)

We'll break this article up into platform-specific lists, which each include our own notes and analysis, and we'll start with the least-surprising list. This collection's stewards at Digital Eclipse (TMNT: The Cowabunga Collection) had already announced plans to "reimagine" six games beyond the original code supplied by Atari's archives, and these alone could be worth the price of admission for '70s and '80s gaming stalwarts.

Two of these, Haunted House and Yars' Revenge Reimagined, were announced as straightforward modern-graphic versions of their Atari 2600 counterparts, and while we trust Digital Eclipse's track record, we have yet to see exactly how these two games have been rebuilt, or whether they lean on classic pixel designs or new 3D assets. Two additional games sound like souped-up, modern remixes of their original counterparts: Quadratank (a four-player touch-up of the 2600's Tank) and Neo Breakout (which combines Breakout and Pong as a two-player versus game).

The last two reimagined games in the collection are the most intriguing. Vctr Sctr will combine "gameplay from a variety of vector arcade classics" into a newly remixed experience. It's unclear whether this new game will break classics up into mini-games or smush them together into a new game experience, but either way, the original Atari arcade library was rich with vector-driven games to choose from. Additionally, Swordquest: Airworld is a new sequel to the 2600's three-part Swordquest series.

This week's leak does not include any previously unknown games in the above "reimagined" category. For surprises, we should move on to the compilation's selection of arcade cabinet games:

• Akka Arrh (1982, unreleased)
• Asteroids (1979)
• Asteroids Deluxe (1980)
• Black Widow (1982)
• Breakout (1976)
• Centipede (1980)
• Cloak & Dagger (1983)
• Crystal Castles (1983)
• Fire Truck (1978)
• Food Fight (1983)
• Gravitar (1982)
• I, Robot (1983)
• Liberator (1982)
• Lunar Lander (1979)
• Major Havoc (1983)
• Maze Invaders (1981)
• Millipede (1982)
• Missile Command (1980)
• Pong (1972)
• Quantum (1982)
• Space Duel (1982)
• Sprint 8 (1977)
• Super Breakout (1978)
• Tempest (1980)
• Warlords (1980)

Since these lists are alphabetical, this selection's biggest surprise is at the top. Akka Arrh is one of the rare few Atari arcade games to be canceled and somehow avoid a wider online release or leak in the years that followed. That changed in 2019 thanks to a controversial dumping of a rare cabinet's ROM chips, so this game's inclusion isn't necessarily a world premiere for MAME enthusiasts. Still, Digital Eclipse and Atari Corp. deserve credit for shaking hands and giving fans a legitimate path toward purchasing and enjoying this unreleased game. (Atari Corp. had previously done this for the unreleased arcade game Maze Invaders, which launched on Atari Flashback Classics Vol. 3 in 2018.)

The collection's biggest concession is made clear by the date range in the above list. Notice that no cabinets beyond the year 1983 appear. That's no accident. Shortly after the early 1980s video game industry crash, the company's corporate masters at Warner opted to break Atari into parts and sell them. The hardware-focused division, Atari Corp., is responsible for this year's anniversary compilation, while Atari Games, Inc., went on to release many incredible Atari-branded arcade games before eventually being acquired by Midway—which, this many years later, is now controlled by Warner Bros. Discovery. (A chance corporate-gaming reunion, then!)

Thus, Atari Corp. and Digital Eclipse apparently have not licensed many notable "Atari" arcade classics from their current license holders at WBD. That list is quite long, but just counting the year 1984, that leaves out the megaton games Paperboy and Marble Madness, with the legendary likes of Gauntlet and 720 following a few years later. Similarly, anything with a significant Hollywood tie-in, like 1983's Star Wars, also doesn't make the cut.

Intel is gearing up to release the first products in its 13th-generation Core processor family, codenamed Raptor Lake. Among the topline facts that the company announced at its Intel Technology Tour is that at least one member of the Raptor Lake family will be capable of hitting 6 GHz out of the box (via Tom's Hardware). Core counts and architectural improvements are generally more important than clock speed when it comes to increasing a CPU's performance these days, but after many years hanging out in the 5 GHz range, it's neat to hit the next digit.

As for what this means for performance, Intel is saying that Raptor Lake will perform roughly 15 percent better in single-threaded tasks and 41 percent better in multi-threaded workloads than the current 12th-generation Alder Lake chips. Clock speed is more important for boosting single-threaded performance, while adding more cores is usually the best way to improve multi-threaded speeds.

It's not clear which of the CPUs will be capable of hitting 6 GHz or under what circumstances or for how long. An Intel SKU chart published by Igor's Lab suggests that the Core i9-13900K will top out at 5.8 GHz, though it's possible that it's capable of further boosting beyond that.

The SKU chart also confirms what we've heard from other leaks—despite an architectural similarity to current Alder Lake CPUs, Intel is boosting performance by doubling the maximum number of E-cores in its top-tier CPUs, from eight to 16. The i5-12600K, which included four E-cores, will also be replaced by a Core i5-13600K that includes eight. And if the rest of that leaked desktop CPU lineup holds true, some lower-end Core i5 processors that shipped with no E-cores in the 12th generation will get four or eight in the 13th generation.

This SKU chart also indicates that these top-end Raptor Lake CPUs will include other changes, including increased amounts of L2 and L3 cache available. It also looks like the Maximum Turbo Power consumption of all the chips is going up by between 12 W and 63 W—from 241 W to 253 W for the Core i9, from 190 W to 253 W for the Core i7, and from 150 W to 181 W for the Core i5. The base power for all the chips stays level, at 125 W.

Those maximum power consumption numbers don't necessarily mean that all Raptor Lake CPUs will consume more power than their Alder Lake counterparts—that will depend on the specific power settings that your motherboard or PC manufacturer chooses to use. It does mean that a Raptor Lake CPU with high power limits and adequate cooling will need more power and better cooling than a similarly configured Alder Lake CPU when running sustained workloads, though.

People old enough to remember installing Windows 98 will recall that clock speed bragging rights were a big deal in the early 2000s. Intel planned to boost its Pentium 4 CPUs from the mid-1 GHz range all the way up to 10 GHz by 2005, but the architecture was just barely able to hit 4 GHz before heat and power consumption both became so high that further increases became untenable (the company's first attempt at a 4 GHz CPU never saw the light of day).

Intel's GPU performance tiers. The gap between 3 and 5 is much larger than the gap between 5 and 7.

Intel’s slow, steady drip of information about its upcoming Arc GPUs continued last week when the company released the final specs for its four A-series desktop graphics cards. As with the laptop GPUs the company announced earlier this year, the desktop Arc cards are split between the entry-level Arc 3, midrange Arc 5, and high-end Arc 7 tiers.

The only Arc GPU that has been extensively vetted by independent testers is the entry-level A380, which is also the only one that has had any kind of official launch in the US (this back-ordered $140 ASRock Challenger model listed on Newegg, not counting more expensive cards imported from China). It's by far the weakest GPU in the lineup, with just eight of Intel's Xe cores and ray tracing units and 6GB of GDDR6 memory on a 96-bit memory bus. Reviews usually show the A380 trading blows with older and budget-focused GPUs like Nvidia's GTX 1650 or AMD's RX 6400, depending on the game.

The other three GPUs are all a lot more similar. The A580, A750, and A770 have 24, 28, and 32 Xe cores, respectively, all increasing in clock speed as you climb upward. But they all share a 256-bit memory bus and 8GB of GDDR6 RAM (with a 16GB option available for the A770). Intel showed off its own "Limited Edition" first-party versions of the A770 and A750; they share the same basic board and fan design and 225 W board power, but the A770 has built-in LEDs, and the A750 doesn't. Intel is presumably leaving the lower-end A580 and A380 cards to its board partners, rather than making its own versions.

Intel's tests show the A750 GPU competing with the Nvidia RTX 3060 in DirectX12 and Vulkan games; going purely from the on-paper specs, this suggests that the A770 will land somewhere between an RTX 3060 and 3070, while the A580 will perform worse than a 3060 but better than a 3050.

This is a pretty narrow range of performance, with a big gap between the low-end and the midrange cards and no offering at all for high-end gaming PCs, but it makes sense given that the Alchemist family is supposedly based on a pair of GPU dies, with one maxing out at eight Xe cores (the A380) and the other at 32 Xe cores (with the 28- and 24-core versions being "binned" versions that shut off defective or underperforming cores so that the die itself can still be used). Nvidia and AMD's current GPU ranges usually use at least three or four different GPU dies.

Making performance comparisons with established Nvidia and AMD GPUs is difficult for Arc GPUs because of driver issues. The company acknowledged that performance in older DirectX11 games wouldn't be up to the same level as performance in games that use the low-overhead DirectX12 and Vulkan APIs, and the company is leaning on code translation for old DirectX9 games rather than allowing the GPUs to run those games natively. And Arc GPU performance suffers significantly if the Resizable BAR (also called ReBar or Smart Access Memory) feature isn't enabled or supported by your motherboard, making Arc GPUs a bad choice for upgrading an older PC.

Compounding Intel's launch issues, AMD and Nvidia are already gearing up for next-generation GPU launches. AMD briefly showed off an RDNA3-based GPU at its Ryzen 7000 announcement event last month, while Nvidia's RTX 4000-series GPUs are expected to be announced as early as next week. Intel’s next-gen Battlemage GPUs won’t be released until 2023 at the earliest, so these Alchemist GPUs will likely find themselves competing against RTX 4060-series and RX 7600-series GPUs soon enough.

Intel has said that it will be pricing Arc GPUs based on its DirectX11 game performance, making its cards more appealing for bargain hunters wanting to put together an inexpensive-but-capable gaming PC. The company just needs to launch them first.

Listing image by Intel

Nvidia announced yesterday that its upcoming H100 "Hopper" Tensor Core GPU set new performance records during its debut in the industry-standard MLPerf benchmarks, delivering results up to 4.5 times faster than the A100, which is currently Nvidia's fastest production AI chip.

The MPerf benchmarks (technically called "MLPerfTM Inference 2.1") measure "inference" workloads, which demonstrate how well a chip can apply a previously trained machine learning model to new data. A group of industry firms known as the MLCommons developed the MLPerf benchmarks in 2018 to deliver a standardized metric for conveying machine learning performance to potential customers.

In particular, the H100 did well in the BERT-Large benchmark, which measures natural language-processing performance using the BERT model developed by Google. Nvidia credits this particular result to the Hopper architecture's Transformer Engine, which specifically accelerates training transformer models. This means that the H100 could accelerate future natural language models similar to OpenAI's GPT-3, which can compose written works in many different styles and hold conversational chats.

Nvidia positions the H100 as a high-end data center GPU chip designed for AI and supercomputer applications such as image recognition, large language models, image synthesis, and more. Analysts expect it to replace the A100 as Nvidia's flagship data center GPU, but it is still in development. US government restrictions imposed last week on exports of the chips to China brought fears that Nvidia might not be able to deliver the H100 by the end of 2022 since part of its development is taking place there.

Nvidia clarified in a second Securities and Exchange Commission filing last week that the US government will allow continued development of the H100 in China, so the project appears back on track for now. According to Nvidia, the H100 will be available "later this year." If the success of the previous generation's A100 chip is any indication, the H100 may power a large variety of groundbreaking AI applications in the years ahead.