AMD Ryzen 9000 review: Impressive efficiency, with bugs and so-so speed boosts

Nearly two years after the release of the first Ryzen 7000 CPUs, AMD has returned with a full-fledged follow-up. The new Ryzen 9000 chips—the 6-core 9600X, 8-core 9700X, 12-core 9900X, and 16-core 9950X—bring AMD's new Zen 5 architecture to the desktop a couple of weeks after it launched in the Ryzen AI chips for laptops.

We came away from the cumbersomely named Ryzen AI 9 HX 370 impressed by both its performance and its power efficiency, and AMD is also leaning into power efficiency as a key selling point of the Ryzen 9000 series. Three of the chips have seen their power limits dropped significantly compared to last-gen chips while still bragging of low-double-digit performance increases. That's rare at a time when Intel has been pushing its chips to the edge to squeeze every last bit of performance out of high-end Core i9 and Core i7 chips.

The focus on power efficiency will give many users—particularly those who don't touch the default settings—less-power-hungry chips that run a bit cooler. And for people who want to tinker and trade in some of that efficiency for a performance boost, many of these chips (particularly the 9700X and 9900X) have a lot of additional performance headroom. It's also nice that all existing AM5 motherboards on the market should be compatible with the 9000 series once a BIOS update is installed, and the AM5 platform's mandatory BIOS Flashback support means you don't need to use weird, kludgy hacks like the "boot kit" CPU loaner program for updating the BIOS when you buy a board with an older BIOS installed.

But it's not all smooth sailing. Motherboard and RAM costs remain an issue across the entire AM5 platform nearly two years after launch, especially for budget buyers. Launch pricing is high, especially compared to current street prices for the 7000 and 7000X3D series CPUs. And we ran into a few bugs during testing that may also bite other early adopters while we wait on AMD and the motherboard makers to straighten things out.

Meet Zen 5

Compared to the Ryzen 7000 launch, when AMD was putting out new CPUs, a new architecture, a new I/O die (IOD), new chipsets, a new integrated GPU, (mandatory) DDR5 support, and a brand-new processor socket all at once, the Ryzen 9000 launch is downright placid. The IOD is the same. The integrated RDNA2-based GPU is the same (and, it's worth reiterating, not at all intended to play games). The AM5 socket is the same. Officially supported DDR5 speeds are up a bit, from DDR5-5200 to DDR5-5600. But the widespread availability of and support for factory-overclocked RAM in desktop motherboards means that a modest bump to officially supported RAM speed doesn't mean all that much; AMD still recommends DDR5-6000 for optimal performance.

There are new 800-series chipsets, but they're so incrementally improved over the old 600-series chipsets that AMD didn't even send along a new motherboard with the new processors. We were directed to install a new BIOS update on existing X670E boards from the Ryzen 7000 launch instead (this almost never happens, even when forward compatibility is a selling point). This seems like a subtle signal from AMD—don't feel like you need to wait for or pay extra money for a newer motherboard, something that would only increase the AM5 platform's stubbornly high base price.

This leaves us with only two new things of note: the Zen 5 architecture itself and the new core chiplet die (CCD) that combines up to eight Zen 5 cores together.

Ryzen 9000's Zen 5 CCD is built using TSMC's N4P process, an upgrade from the 5 nm process used for the Ryzen 7000 series. TSMC says that N4P is 22 percent more power efficient than the 5 nm process, which is part of what is enabling AMD to use a lower TDP for most of the Ryzen 9000X CPUs while still providing faster performance.

It also probably has at least a bit to do with the lower operating temperatures of Ryzen 9000, which AMD has said should be about 7° Celsius lower than equivalent Ryzen 7000 CPUs at the same power levels. Ryzen 7000X chips running at their default TDPs ran a bit hot, and while AMD said this was normal and safe behavior, a cooler chip will usually offer better consistency and longevity.

A major architectural change from Zen 4 relates to support for AVX512 extensions, which Intel introduced (infamously, in some circles) in 2017 with some of the higher-end Skylake variants but hasn't enabled on recent consumer CPUs because Intel's E-cores don't support them. Zen 4 brought AVX512 instructions to Ryzen CPUs for the first time, but it did so by combining two 256-bit data paths. Zen 5 has a full 512-bit data path for AVX512 instructions, and for the (still rare-ish) tasks that actually use AVX512, it gives Zen 5 a disproportionately high performance boost.

There's one thing Ryzen 9000 doesn't include: There's no integrated neural processing unit (NPU) here. So far, AMD has only included those in laptop processors—like the Ryzen 7040, 8040, or Ryzen AI 300—or in CPUs like the Ryzen 8000G series that repackage laptop silicon for use in desktop motherboards. If you're using a desktop for AI-related work, there's a good chance you're using a powerful dedicated GPU for it. But it leaves regular desktop users without any option for supporting Copilot+ or other NPU-accelerated features. To be clear, NPUs and Copilot+ are still in their infancy in desktop PCs, and even when they're present, they don't do all that much yet. It's just odd to hear AMD talk up its XDNA technology in laptops while ignoring it in high-performance desktops.

Testbed notes

The core of our CPU testbed is the same as it's been for a while now: a Lian Li O11 Air Mini case with an EVGA-provided Supernova 850 P6 power supply and a 280 mm Corsair iCue H115i Elite Capellix AIO cooler. We use an AMD-provided Radeon RX 7900 XTX as our GPU, so most benchmarks running at 1080p will be more CPU-bottlenecked than GPU-bottlenecked.

As usual for our CPU reviews, we also try to test each model with different power limit settings in addition to the manufacturer-specified defaults. For most Intel and AMD chips, this is an easy and relatively foolproof way to either boost performance at the expense of power efficiency or lower performance while improving efficiency (AMD even supports this directly with its "Eco Mode" settings, which automatically sets each CPU's TDP setting one level lower than it is by default).

For Intel's chips, these settings are called PL1 and PL2 (Power Level 1 controls how much power the CPU can use during a sustained, intensive workload, while PL2 dictates how much it can use for short bursts). For AMD boards, TDC is roughly equivalent to PL1 and EDC is roughly equivalent to PL2, while a third number (PPT) determines the amount of power provided to the CPU socket.

Testing difficulties

When AMD and Intel send us new CPUs to test, they provide a reviewer's guide with some background information on the chips. Included in these guides are some manufacturer-provided reference numbers for some common, easily reproduced benchmark tests—not so reviewers can just republish them but so we can check to make sure our test beds are performing more or less as expected when we go to run our own tests.

Normally, this is pretty trivial. AMD and Intel also provide motherboards, RAM kits, pre-release BIOS updates, and pre-release drivers, so there really aren't many unknown variables that should prevent the chips from performing as expected; usually, we can just plunk them in, set an XMP or EXPO RAM overclocking profile, run some perfunctory tests for reference, and get on with it.

The 9600X and 9700X in particular weren't performing as expected and were performing worse with memory overclocking profiles applied—often no better than the Ryzen 7000-series CPUs they're supposed to be replacing. Other reviewers, such as the YouTubers at Gamers Nexus and JayzTwoCents, also reported some stability and RAM configuration problems.

After much discussion and sharing of log files with AMD—as well as multiple testing runs with both processors, three different motherboards, and three different RAM kits—the issue seems to be that the EXPO memory overclocking profile is applying incorrect settings. Specifically, we had to set the SoC voltage to 1.2 V from the default 1.25 V, and we had to set the Infinity Fabric clock speed (or FCLK) from 2100 MHz to 2000 MHz. There may be other things that could be tweaked, but this got our numbers to where they were supposed to be.

AMD told us that part of the issue came down to that 1.25 V voltage setting interacting with the 65 W TDP for both the 9600X and 9700X—that with a relatively limited power budget, delivering more power than necessary to the CPU SoC takes it away from the rest of the processor. This didn't explain why the EXPO profile was apparently loading the wrong SoC voltage in the first place, but it does at least make logical sense.

The upshot is that any early adopters who attempt one-click overclocking will probably run into the same issues we did. These are hopefully things that can be cleared up with future BIOS updates, but if you're buying these chips on day one, you may have some problems getting the full performance out of them. AMD didn't give us any kind of timeline for a fix.

Performance

For simplicity's sake, we've compared the Ryzen 9000X series primarily to the Ryzen 7000X series since the new chips are direct replacements for the old ones, and the non-X Ryzen 7000 series aren't really any cheaper than the 7000X chips at this point. We've also included comparisons to the 7000X3D series, which are cheap enough that they compete with the 9000X chips on price. A smattering of Ryzen 5000-series CPUs are also included to provide data to people with a 3- or 4-year-old PC they're thinking of upgrading.

On the Intel side, we're comparing Ryzen 9000 mostly to 13th- and 14th-generation Intel chips, though note that all the Intel performance numbers here were taken before we knew anything about the voltage issues that have caused some of those processors to hang and crash. Intel has said that the microcode update can reduce performance slightly under some circumstances, though in general, there shouldn't be a noticeable difference.

AMD generally delivers low-double-digit improvements to single-threaded performance across all of the Ryzen 9000 chips and pretty similar single-core performance regardless of whether you go with the 9600X or the 9950X. Multicore improvements are more all over the place, at least in our testing; the 9600X is usually mid to high single digits faster than the 7600X, and the 9950X is usually 10 or 11 percent faster than the 7750X in our tests. But the 9700X sometimes has trouble outperforming the 7700X at all. That's still arguably a win—the 9700X has a default TDP of 65 W instead of the 105 W TDP of the 9600X. But you're definitely trading speed for efficiency here.

The higher-end 13th and 14th-generation Core chips from Intel manage to keep pace with the 9950X's multi-core performance, mostly thanks to the sheer number of cores, an advantage of Intel's heterogenous approach to its desktop processors. But there are outliers in either direction; 3DMark Time Spy and Shadow of the Tomb Raider both run quite a bit faster on Intel's cores, but our H.265 video encoding benchmark is a lot faster on the Ryzen 9000 chips thanks to their beefed-up AVX512 support (the 9000-series CPUs are disproportionately faster than the 7000-series CPUs here, too).

This is commonly understood among more veteran desktop builders, but look at the game benchmarks we've included here and remember that for a gaming-focused PC, there's still very little reason to buy anything faster than a good Ryzen 5, Core i5, Ryzen 7, or Core i7 CPU. Plowing more money into a CPU gets you almost nothing in terms of game performance, even at low resolutions—spend more money on your GPU instead. Alternatively, buy one of the Ryzen 7000X3D chips, which still often beat the Ryzen 9000 CPUs in gaming benchmarks.

Power efficiency

For the Ryzen 9600X, 9700X, and 9900X, all three chips generally manage to outperform their predecessors despite being set to lower TDP levels and consuming less power.

We do see indications that this is holding the 9000-series back. Setting the 9700X to a 105W TDP helps it leap ahead of the 7700X, though its power consumption leaps ahead quite a bit as well. Using AMD's Precision Boost Overdrive should yield good results here since it's meant to help find the best spot on the performance/power curve for specific CPUs.

At its stock settings, the 9950X is the one exception in our testing. While both the 9950X and 7950X have a 170 W TDP, the 9950X actually pushes up against the limit of that TDP number, performing better but also burning more power when all the CPU cores are engaged in heavy-duty tasks like video transcoding. (It is, however, more efficient than Ryzen 7950X by a bit when both chips are set to their 105 W Eco Mode TDP).

When experimenting with higher TDPs for some of the Ryzen chips, we found that each processor will gladly run hotter and consume more power when allowed to. The 9700X is a good example—when manually set to a 105 W TDP, you can expect around 10 percent faster performance in heavy multithreaded tasks, though the CPU can consume as much as 60 percent more power when all its cores are active at once.

The Intel comparison is much more flattering for AMD. Whether running at its stock 170 W TDP or in its 105 W TDP Eco Mode profile, the 7950X transcoded our test video as fast or faster than the Core i9-14900K while using a fraction of the power. The 9600X and 9700X are a bit less convincing compared to the i5-14600K and i7-14700K—Intel's core count advantage is still tough for AMD to overcome, here—but AMD still delivers great game performance and good power efficiency even in these workloads.

Conclusions

Right now, Ryzen 9000 chips are a tough sell. While they are improvements over the existing Ryzen 7000-series, they don't deliver the same generational leap we saw between Zen and Zen 2, Zen 2 and Zen 3, or Zen 3 and Zen 4. Ryzen 9000 is certainly more efficient. But Ryzen 7000 was already reasonably efficient, particularly in Eco Mode. And as of this writing, the Ryzen 7800X3D isn't even $20 more expensive than the 9700X, and it comes with some of its own performance advantages.

On the other hand, the regular Ryzen 7 7700X and Ryzen 5 7600X (both currently available for about the same price as the non-X variants) are 70 percent and 80 percent of the price of their Ryzen 9000 counterparts while delivering about 92 percent and 98 percent of the speed (depending heavily on the specific app or game and your power settings, of course). Socket AM5 still wouldn't be my first choice for a budget build, but at the moment, Ryzen 7000 and Zen 4 are delivering a better value.

That's all here and now, though. Taking a longer view, assuming that new Ryzen 7000 chips will eventually vanish and Ryzen 9000 chips will slide downward a bit in price, they're respectable upgrades. We're always big fans of chips that balance speed and efficiency, something AMD is doing a better job of than Intel these days. If you're sitting on a Ryzen 3000 or Ryzen 5000 system you bought, built, or upgraded three or four years ago (or an Intel system from the 2015–2020 Skylake Forever era), these will give you that much more speed if and when you decide to upgrade your system.

Still, it's hard not to feel a touch of disappointment when using Ryzen 9000. Early rumors suggested the desktop chips might copy Intel's or Apple's recent desktop CPUs, swapping to a heterogeneous design that boosted single-core performance with a few large fast cores while supercharging multi-core performance with gobs of smaller efficiency cores. AMD is, in fact, taking this approach with the Ryzen AI 300 series laptop chips, combining four Zen 5 cores with eight smaller, slower Zen 5C cores. AMD's desktop core counts have remained the same since 2019, when the 16-core Ryzen 3950X added a new high-end option to AMD's consumer desktop platform.

If AMD plans to take that route for its desktop chips eventually, it isn't doing so with the first wave of Ryzen 9000 CPUs. There's always next generation.