AMD Ryzen 9 9900X and Ryzen 9 9950X review: mystifying performance

A week ago we looked at the slightly underwhelming AMD Ryzen 9600X and 9700X, which were marked by modest gaming performance increases in some titles, more significantly better content creation grunt and slightly better thermals. Now it’s time to look at the second half of AMD’s Zen 5 quartet, the £459/$499 Ryzen 9 9900X and £609/$649 Ryzen 9 9950X. These are powerful 12-core and 16-core parts that ought to be more of interest to content creators than gamers, so do they make a better case for Ryzen 9000? And does either one manage to exceed the top-level gaming performance of the 7800X3D?

Unfortunately, after four days of frantic benchmarking and troubleshooting, I’m not sure AMD has succeeded on either of these points. Our Ryzen 9900X and 9950X testing has been marked by confusingly poor gaming performance, including performance regressions versus the 7900X and 7950X, alongside a few examples of genuine uplifts that nonetheless don’t go far enough to making these CPUs worth recommending.

This review does at least mark the debut of two improvements to our benchmarking suite compared to last week’s reviews, which itself marked the introduction of an entirely new automated benchmarking system for Digital Foundry – and therefore only included a more limited selection of games than we’ve offered in the past.

First, we’ve added 720p data – in addition to the 1080p, 1440p and 4K data – which some people prefer for CPU benchmarking as it takes the GPU even more out of the equation. Second, we’ve added two new games in new genres: Counter-Strike 2 to represent competitive shooters and Starfield to represent Bethesda-style RPGs. Both offer a stern challenge for modern CPUs, especially with the higher refresh rate screens now common. Of course, we’ll continue to expand our testing in the future, with some exciting possibilities on the horizon.

Based on our 9600X and 9700X results last week, while gaming performance doesn’t match the pace of improvement in previous Ryzen generations, we do at least expect to see a reasonable uptick in content creation performance. That’s thanks to an increase in IPC (instructions per clock, an indicator of single-core grunt), while clockspeeds, power usage and thermals remain flat or slightly improved. As with the 9600X and 9700X, the single-core gains come as a combination of architectural improvements (such as an improved branch predictor, better AVX-512 support and doubled data bandwidth between L2 to L1 caches) and the switch to a more efficient 4nm CCD.

With that in mind, the 9900X is rated at a lower 120W TDP than the 170W 7900X – though the 9950X is at the same 170W mark as the 7950X it replaces. Improved thermal resistance also means that temperatures at the same TDP ought to be lower – AMD claims a 7°C reduction at matched TDPs, which is arguably more important in these high core count products more often used for all-core workloads that maximise temperatures.

As suggested by AMD, we’re using a similar physical setup to our previous Ryzen 7000 testing, including an ASRock X670E Taichi motherboard, G.Skill Trident Z5 Neo DDR5-6000 CL30 RAM and an Eisbaer Aurora 240mm AiO. The big difference is that our RTX 3090 has been swapped out for an RTX 4090 Founders Edition, providing extra graphics performance to push modern CPUs even harder.

While Ryzen 9000 is the star of the show, we also tested a selection of earlier Ryzen 3000 and 5000 processors and a selection of Intel’s 14th-gen processors. Our Intel CPU testing was performed on the Gigabyte Aorus Z790 Master using the same DDR5-6000 CL30 RAM, while AM4 CPUs were tested on the classic Asus ROG Crosshair 8 Hero with a Trident Z Royal DDR4-3600 CL16 kit.

For storage, we’re using a 4TB Lexar NM790 PCIe 4.0 NVMe SSD and 4TB Kingston KC3000. Our rig is completed with a 1000W Corsair RM1000x power supply. Testing was performed on a fresh install of Windows 11 with the latest Windows updates (23H2), chipset drivers (6.06.28.910) and BIOS revisions (3.06 for the ASRock AM5 board) installed.

Before we get into the gaming benchmarks, it’s worth spending some time looking at the results of our content creation testing. These serve as synthetic benchmarks for gamers, suggesting certain levels of performance we might expect in a best-case scenario, and of course they can also be a useful indicator for people that actually intend to use a Ryzen 9000 CPU for 3D modelling or video production. These 12-core and 16-core parts tend to be impressive performers, and thankfully that is the case for Zen 5 as well.

We’ll look to expand this in future, but for now we have three workloads: Cinebench 2024 and R20 simulate the 3D modelling and animation program Cinema 4D, while Handbrake is literally just transcoding an early DF Patreon video into H264 and H265 formats – a common task for anyone that works with video.

Let’s look at the Cinebench 2024 results first, which are recorded after a 10 minute plus loop to ensure CPUs aren’t able to quickly knock in a good score before thermal or power limits kick in. Here we see single-core improvements in the ~15 percent range for both CPUs over their predecessors, while multi-core gains are a bit more modest: 11 percent for both. For context, Intel’s 14700K is 11 percent faster than the 9900X in this test, but the 9950X outduels the 14900K by around six percent.

Cinebench R20 is a legacy test for us, having run in a huge number of previous CPU reviews, and so we thought it was worth running as well in case you wanted to go back and compare results. We saw similar margins here gen-on-gen (+11 percent multi-core for both Ryzen 9000 CPUs).

Our final test is the Handbrake transcode test, where we convert an 822MB 4K video file using the H264 and H265 codecs using the Production Standard preset set to CRF 18. We’re using the most recent version of Handbrake here, 1.8.1, which helpfully shows its statistics in the application window rather than burying the average encoding frame-rate in a log file.

These transcode results show a much smaller advantage for Ryzen 9000, something we also found with the 9600X and 9700X. That’s around five percent in H264 and eight percent in H265 for the 9900X and 9950X over their immediate predecessors – perhaps due to those AVX architectural improvements.

We didn’t have the chance to experiment with PBO and other performance enhancements this time around, but we did see a decent jump at the expense of power efficiency with the 9700X and I’d expect that the 9900X and 9950X will get meaningfully faster if you disregard the stock power and thermal budget.

With our content creation benchmarks completed, it’s time for the game benchmarks. Unfortunately, it’s largely downhill from here.

AMD Ryzen 9 9900X and Ryzen 9 9950X analysis