AMD Ryzen 9 9950X in detail
The most powerful AMD Zen 5 CPU is the fastest not only multi-threaded (when all cores are involved), but also single-threaded. It usually beats the competing Core i9-14900K processor quite clearly. Although, this may not always be the case and there are situations where the 16-core Ryzen 9 9950X is weaker. I mean, speed-wise. The efficiency (due to the lower power consumption) may already be on its side.
Conclusion
The new Ryzen 9 9950X can be considered the fastest desktop processor available today, excluding Threadrippers. It outperforms its predecessor (R9 7950X) by 12–14% in heavy 3D rendering workloads. The lead over the Core i9-14900K is already a little smaller – 5% – but this with 16% lower power consumption. This makes the Ryzen 9 9950X 27% more efficient. However, the R9 9950X already trails the R9 7950X by 4% in efficiency, although the newer model is faster. When encoding x264 video in Handbrake, the difference in efficiency (in favor of the Ryzen 9 9950X) is already greater – 10%.
The top-notch performance for x264/x265 encoding is also well suited for video editing programs such as Adobe Premiere Pro or DaVinci Resolve Studio. The Ryzen 9 9950X also made its way to the top spot in Topaz Labs AI photo restoration apps. Compared to the competing Core i9-14900K, the AMD processor is as much as two and a half times faster. It owes this to the very efficient implementation of the AVX-512 instruction set with VNNI support.
The R9 9950X also ranks at the top in Affinity Photo tests, in Zoner Photo Studio X (converting uncompressed photos to JPEGs). For batch editing in XnViewMP, the Intel CPU (Ci9-14900K) is still faster, which usually has the upper hand in Adobe Photoshop as well. Exporting RAWs or generating thumbnails (1:1) in Lightroom is already quite even.
(De)encryption in TrueCrypt is already dominated again by the Ryzen 9, where the Core i9-14900K sometimes runs at half the speed. This is the case, for example, in the Serpent tests. This is probably related to the fact that the L1 and L2 caches of Zen 5 cores have twice the data width and thus twice the theoretical throughput, which is confirmed by tests in Aida64. The benefit from this may not always come, in fact it is quite rare, but sometimes it does show up. Performance during (de)compression is like a seesaw, sometimes the R9 9950X has a bit of an edge, other times the Ci9-14900K does.
The gaming performance is stagnant. It can be a hair higher in one game only to be a hair lower in another game. But the average is comparable, with the caveat that if you’re after even the smallest differences, you’ll probably get better results in terms of framerate with the Core i9-14900K in most cases. In terms of efficiency, the latest Ryzen 9 – Granite Ridge – is the best of the trio (Ci9-14900K and R9 7950X and R9 9950X). Its power consumption in games is the lowest, also shaped by the lowest all-core boost clock speeds.
But in heavy multi-threaded workloads, the Ryzen 9 9950X already consumes more power than its predecessor, unless you tone down its power supply. The increase here is at about 30 W, with which it’s still some 50 W below the Core i9-14900K.
However, the power consumption in a single-threaded workload is already higher compared to the Ci9-14900K, even compared to the Ryzen 9 7950X, by about 20%. Here, however, it’s good to know that single-core boost clock speeds are heavily dependent on the temperature. And since that’s been reduced, or at least lower temperatures are being reported, there’s more headroom than there used to be (with the R9 7950X). On average, with the same cooling performance (of the Noctua NH-D14S cooler), we are thus at 150 MHz higher SC boost clock speeds (approx. at 5650 MHz). From these, the elite performance in single-threaded applications is then derived. Be it encoding of audio recordings, the responsiveness of the web environment (Google Chrome) or work in office applications. These are tasks where it can be noted that the Ryzen 9 9950X currently has no competition. The single-threaded performance is very high indeed, albeit at the expense of relatively higher power consumption.
Some more findings on temperature: remarkably, the processor stays below 95 °C even under very heavy load. This creates better room for tuning coolers compared to the last generation and the possibility of using cheaper models. Although, there probably aren’t many reasons to skimp on a cooler with a processor like the Ryzen 9 9950X. The more powerful the better. It’s just that with this processor there is more margin for cooling optimizations, for example, with a view to running a quieter high-performance PC build.
English translation and edit by Jozef Dudáš
| AMD Ryzen 9 9950X |
| + Extreme multi-threaded performance surpassing the Core i9-14900K... |
| + ... at a significantly higher efficiency |
| + Highest single-threaded performance we've ever seen in tests |
| + Top-notch gaming performance |
| + "Versatile" processor, fits every usage scenario |
| + 16 cores and 32 threads on a mainstream platform |
| + Very high performance per clock (IPC) |
| + State-of-the-art 4nm manufacturing node |
| + Very high clock speeds given the new manufacturing node |
| + DisplayPort 2.0 support |
| + Favorable price/performance ratio for a high-end CPU |
| + Lower temperatures than previously (with Ryzen 9 7000)... |
| - ... temperature-wise the processor still lands high and a more powerful cooler is suitable |
| - Need for a very powerful cooler to achieve maximum single-threaded performance |
| Approximate retail price: 649 EUR |
We would like to thank the Datacomp e-shop for their cooperation in providing the tested hardware
Special thanks also to Blackmagic Design (for DaVinci Resolve Studio license), Topaz Labs (for licenses to DeNoise AI, Gigapixel AI and Sharpen AI) and Zoner (for Photo Studio X license)
- Contents
- AMD Ryzen 9 9950X in detail
- Methodology: performance tests
- Methodology: how we measure power draw
- Methodology: temperature and clock speed tests
- Test setup
- 3DMark
- Assassin’s Creed: Valhalla
- Borderlands 3
- Counter-Strike: GO
- Cyberpunk 2077
- DOOM Eternal
- F1 2020
- Metro Exodus
- Microsoft Flight Simulator
- Shadow of the Tomb Raider
- Total War Saga: Troy
- Overall gaming performance
- Gaming performance per euro
- PCMark and Geekbench
- Web performance
- 3D rendering: Cinebench, Blender, ...
- Video 1/2: Adobe Premiere Pro
- Video 2/2: DaVinci Resolve Studio
- Graphics effects: Adobe After Effects
- Video encoding
- Audio encoding
- Broadcasting (OBS and Xsplit)
- Photos 1/2: Adobe Photoshop and Lightroom
- Photos 2/2: Affinity Photo, Topaz Labs AI Apps, ZPS X, ...
- (De)compression
- (De)encryption
- Numerical computing
- Simulations
- Memory and cache tests
- Processor power draw curve
- Average processor power draw
- Performance per watt
- Achieved CPU clock speed
- CPU temperature
- Conclusion
Thanks for the test ! Suggestion : could you add some specific audio benchmark like DAWBench ? Those are kind of hard to find it could be useful for some people (not a lot obviously but the audio creator community would certainly be glad to have a reliable source).
Thanks for the tip for a test. Sure, we can add DAWBench to the test methodology, but I will collect data in it only for processors not yet tested, i.e. from R7 9700X > R9 9900X > Intel Arrow Lake > …
It looks like you have mistakenly transposed the values for the 7800X3D in Fire Strike’s Graphics score and Fire Strike’s Combined score on page six.
A few other test anomalies(?):
On page 31, the AIDA64 AES / SHA3 test shows both the 9950X and 9600X as severely underperforming. It would be interesting to know why, whether a software issue or otherwise.
On page 32, for the Convolution (SPECworkstation 3.1) test, the Core i5-12400 (C0) seems to have been given a strong energy drink, and it outscores all of the other CPUs by a factor of nearly 4x. Seems like an error. 🙂
On page 33, the 9950X (and 9600X) are missing from the RodiniaCFG test charts, but I did not see a note explaining why this was the case (as was present on the AIDA L3 blank result issue).
Power and temperatures were also both higher than most other reviews, but it may be explained by variance in the power measuring device and use of an air cooler rather than a suggested AIO.
Good review, it’s perhaps not an exciting processor, but is an interesting one. Really excels in certain areas but fails to move the bar in others. I suspect that there are a small minority who are very excited about it for their specific workloads (AI/ML/encryption). Cheers!
The poor AES/SHA3 Aida64 results are probably due to unfinished SW optimizations (on the Aida64 side). This is not a mistake / incorrect description of the results.
Yes, the Convolution value for Core i5-12400 (C0) is indeed wrong. It was an accidental mistake to describe “score” instead of “Spec ratio”. Corrected. Thanks for the heads up!
RodiniaCFG test results for some processors are missing because they end up with an error. We’ve added a traditional explanatory note at the end of the chapter to make everything clear.
Power consumption: Many testers are satisfied with the motherboard’s SW reporting and rely on the “CPU power package”. In my experience, these values are often inaccurate and the power consumption ratios between processors often do not correspond to reality. This is despite the fact that these values are based on the processor’s own management and maybe that’s why they are sometimes understated. Our measurements are based on hardware, we measure the electrical current on EPS cables. A certain disadvantage of these measurements is that they also reflect the (in)efficiency of the motherboard VRM, so that is why the measured values are higher than the real consumption of the processors. But always more or less equivalent. We strive for a comparatively high VRM efficiency across the test motherboards. Nevertheless, of course, there are differences between the motherboards. That’s why we also cover the motherboard tests in detail, in which you can see that for example, with a comparable performance of the R9 7950X you measure 215.7 W (ASRock B650E Taichi), but also 258.9 W (MSI B650 Gaming Plus WiFi). Yet the power limits are set the same.
Higher temperatures than in other tests are probably due to the relatively lower cooling performance of the Noctua NH-U14S compared to for example the more powerful 360 mm AIOs that are commonly used in tests?
—“…due to the relatively lower cooling performance of the Noctua NH-U14S…”
Since you mention the NH-U14S, are you going to compare the standard mount vs offset vs offset + NF-A15 Gen2 anytime soon? 😛
We don’t have it in the roadmap of tests yet. And although it would certainly be interesting (to test an “old” heatsink with a new fan), the list of things to test is extremely long… and, unfortunately, not everything can be done. But what we will definitely be testing is the new Noctua single-tower cooler with a 140 mm fan. No, I don’t know that Noctua is planning one, but I believe that its addition to the cooler lineup will only be a matter of time. 🙂
—“… what we will definitely be testing is the new Noctua single-tower cooler with a 140 mm fan…”
Well, I’m guessing a time horizon of at least 2 years (I’m an optimist 🙂 )…
And in that case, will you also test the old (gen1) A15 fans on the new cooler?
… although at that time it already… 😛
We’ll see, I don’t want to make any promises in this direction. What we are going to focus on are tests of the NH-D15 G2 on different platforms. Soon we will start on LGA 1700 and then, as part of the new, long-term methodology, we will also test on AM5 (9950X) and LGA 1851. But I understand that the tests of the NH-U14S with the NF-A14x25r G2 would be more precious, since few (if any :)) will be dedicated to them and maybe we will eventually do them if the wait for the successor is too long. 🙂
—“…maybe we will eventually do them…”
😘
Appreciate the followup. Thank you!
Thanks Brian, good point! Indeed the results of the graphical score were switched with the combined score. Fixed (also in older tests)! 🙂