Due to high demand, we are investigating the performance of AMD's Threadripper 2990WX on Linux. Apparently, according to many AMD fans, this is something we should have done since day one, although of course this is not part of our usual test battery.
If you missed it, we conducted a comprehensive test of the Ryzen Threadripper 2990WX and Threadripper 2950X processors on the first day. The latter is the replacement for the previous flagship Threadripper and overall an excellent service provider for the price. However, the 2990WX is a different kind of animal, more expensive, and offers more cores, but with mixed results. For this reason, we ran the Mega Tasking Test, a separate, multitasking benchmark to examine the workstation performance of these high-end CPUs in more detail.
Today we used the Threadripper 2990WX under Windows 10 and Linux, especially Ubuntu 18.04.1 LTS, directly against the Core i9-7980XE. Both operating systems were tested in their standard configuration, so that no optimizations were made.
Both test systems have been configured with 64 GB DDR4 memory, the 2990WX is limited to 3000 MHz, while the 7980XE would be happy to accept 3200 MHz and go higher, but have to try to keep things from apples to apples as much as possible we chose CL14 DDR4-3200 memory with low latency.
M-queens is part of the Phoronix test suite and it was one of the few benchmarks in the suite that we were able to work with under Linux and Windows as well as AMD and Intel hardware. The test suite was a nightmare, to be honest.
The benchmark measures the time to solve the problem of the n-queens. M-Queen only uses a much larger board, which makes it a longer, more complex problem. Here we measure the time that needs to be resolved, and as you can see, the 2990WX does it in about half the time of the 7980XE on Windows or Linux, which is extremely impressive.
Stockfish is a free open source chess engine that is available on various platforms. It is consistently ranked as one of the best chess engines and is the most powerful open source chess engine in the world. The speed of this test is measured in node positions per second, essentially positions per second, and this is entirely determined by the performance of the processor.
Again we see, when it comes to chess, the 2990WX is a beast on Windows or Linux. This time, however, we see a 23% increase in performance for the 32-core processor under Linux. Meanwhile, the 7980XE only saw a 4% increase in performance when tested on Linux.
The last chess benchmark we're going to look at is Crafty. Again, we measure performance in knots per second. Interestingly, the Core i9-7980XE wins here and has seen the greatest performance increase when switching to Linux. The 2990WX saw an increase in performance of 5% compared to only 3%, which made the Intel CPU 12% faster overall.
John the Ripper is a password cracker that is currently available on multiple platforms, including Linux and Windows. Using the traditional data encryption standard, we see that the 2990WX and 7980XE execute approximately the same number of crypts per second on Windows. On Linux, however, it is a completely different ball game. Here the 7980XE is a whopping 80% faster, but the 2990WX is really impressive and offers three times more power. It is about 80% faster than the Core i9 processor.
We find a similar story when we use the Blowfish cipher. Here the 7980XE is 7% faster on Windows, while the 2990WX is almost 70% faster on Linux. While the 32-core processors look less than impressive on Windows 10, it looks incredible on Linux.
GraphicsMagick is a simple but highly efficient software package for viewing and editing images. It supports a variety of formats and image processing is heavily multithreaded. It is used by several websites to process a large number of uploaded images. Here we look at the sharpness performance measured in iterations per minute, so higher is better.
On Windows 10, the 7980XE is 26% faster than the 2990WX. This changes on Linux, however, and now the 32-core processor is 10% faster. Not a huge win, but the AMD processor comes from behind on Windows to beat the Core i9 on Linux.
The improved operation offers 30% better performance on Linux, but we're seeing similar gains for both AMD and Intel processors. The end result is a small win for AMD.
Next we have the Apache benchmark, which measures the performance of HTTP web servers and as you can see, the performance on Windows is extremely poor. Windows Server is a bit better, but is still smashed by Linux, which is why most web servers run Linux (e.g. Catrachadas uses nginx on Linux boxes instead of Apache). On Windows 10, the 7980XE was 4% faster than the 2990WX, but on Linux the 18-core processor was almost 40% faster, which isn't a great result for the 2990WX here.
We saw earlier that coding performance is not particularly good with the 2990WX, and we see again that the 7980XE is 6% faster on Windows. We see almost the same thing on Linux. The 7980XE can displace the 32-core processor. For encoding workloads, Linux seems to be doing little to help the 2990WX.
The 2990WX has also reached our VeraCrypt benchmark and we find some strange results on Linux. For the 50 MB test, the 2990WX under Windows 10 was slightly slower than the 7980XE. Then the performance of the Linux tanks, especially for the 7980XE.
But where the 2990WX really had problems on Windows was the memory-intensive 1 GB test. Here it offered half the performance of the 7980XE. This result was improved on Linux, but the 7980XE was still 35% faster. The 1 GB results are roughly in line with our expectations. However, we are not sure what is going on with the 50MB results on Linux.
We also found some strange results with 7-zip on Linux. The exceptional decompression performance of the 2990WX has been reduced by 13% when using Linux as opposed to Windows 10. The performance of the 7980XE also decreased, although it was very low.
The compression performance of the 2990WX has been dramatically improved. Here we see a massive performance increase of 54%. That said, the 7980XE also saw a performance boost, although it was 10% less. However, this meant that the 7980XE was still 11% faster overall than the 2990WX.
Finally, we will run the Blender Open Data benchmark suite and expect positive results for the 2990WX as it has already proven to be a rendering powerhouse. Here we have the Barbershop Interior Test and measure the completion time in seconds.
The 7980XE was 23% faster on Linux, while the 2990WX was 31% faster. This meant that the 2990WX was 21% faster on Windows than the 7980XE, and 29% faster on Linux.
We see a slightly different scenario in the BMW27 test, although the 2990WX is still far ahead in the result. Since the 32-core processors were 52% faster under Windows 10, this margin was reduced to 45% under Linux.
This time we see a similar scaling with both CPUs. Both were 48% faster in Linux for classroom workload.
This time Linux preferred the 2990WX slightly and made it 59% faster, while it was 52% faster on Windows.
We see a similar story with the Koro workload: Linux slightly improves the 2990WX's position, but overall, both CPUs see good gains on Linux.
Then we have the workload of the Barcelona Pavilion and we see more of it, so time to summarize it.
Does Windows 10 affect Threadripper 2990WX performance? Undoubtedly yes, but it also chimps the 7980XE to a certain extent. If you only compared the 2990WX on Linux and Windows 10, the situation would be much more extreme than it really is. For example, you would conclude that the 32-core processor will ruin the 7980XE even more if both CPUs actually show an almost identical increase in performance.
We still deliver strange results under Windows programs like 7-zip and VeraCrypt. For example, the compression performance of the 2990WX has been improved, but the decompression performance has been weakened. Overall, however, Linux offered a better balance.
Although we could still see the effects of the multi-die design on the 2990WX and its "limited" memory bandwidth, to put it in any way in the 7-Zip compression test, even on Linux and jumps in performance throughout the test, but average was very similar after each run.
The coding performance also remains poor, although we have only carried out very limited tests on Linux so far. For non-memory-sensitive workloads like rendering Blender, the 2990WX is a beast here. This also includes the benchmarks for chess and password cracking.
Depending on what you want to use the 2990WX for, it will determine how useful it really is. We believe that Windows 10 optimizations are possible to better use or better manage these core-heavy CPUs. Improved thread planning doesn't make the 2990WX a weapon in workloads where it is difficult, but we hope that under these conditions it can at least keep up with the 2950X instead of slowing down a lot.
Results like what we saw in the GraphicsMagik focus test were promising as the 2990WX went from much slower than the 7980XE on Windows to slightly faster on Linux and we saw the same thing on John The Ripper. However, this is only a small selection of applications and does not really benefit the content creators.
We'd like to test Premiere on Linux to see if our custom warp stabilizer test, which runs a dozen instances at a time, works better. On Windows, this test maximizes the 2990WX, but the resulting performance is very poor.
In conclusion, we have to say that the few days we spent with Linux were interesting, but ultimately we appreciated the experience with the Windows 10 desktop operating system.