Epyc 9575F vs. TR 9985WX performance discrepancies?

Hardware Hub
,
1

Hi!

In looking at the 9575F compared to the 9985WX, I expect the performance to be fairly similar based on the specs, when memory bandwidth is not critical. The 9575F has a 3.3GHz base / 5.0GHz boost with 400W power limit, while the 9985WX has a 3.2GHz base and 5.4GHz boost with 350W power limit. For tasks dependent on single-thread performance, the 9575F seems like it should be about 93% as fast.

However, I’ve seen a few tests where 9575F appears to suspiciously underperform. Test setup issues are most likely.

StorageReview test of 9575F: only 1,052 points in Cinebench R23 single thread.

Chinese eBay sellers listing 9575F QS CPUs (many of them reusing images): 1657 in Cinebench R23 single thread.
QS results may not be reliable anyway, but still notable.

9985WX should get about 2150, so I expect around 1990 for 9575F.

In these tests I also notice that the performance in the “Index of Nearest” test is only half of TR 9980X, which is significant. The “Mesh Fracturing” test is also worse, although in that case it almost looks like it’s scaling negatively with more memory channels as 9995WX is also slower than 9980X and 9970X but faster than 9575F.

Some small penalty from memory latency seems like it would be normal, but these look like too much.

How should this be interpreted? If these are testing issues, would anyone with a 9575F be at least willing to run Cinebench R23 to see what happens? If this is normal behavior, why?

I should also add that if anyone is running a 9575F + Windows, I would be really interested in seeing some basic gaming benchmarks. Absurd, you say? Plenty of game benchmarking has been done on Threadripper, which isn’t really any crazier. It’s useful for evaluating performance in game development and related work. If certain tasks are unexpectedly much slower than expected, that’s useful information. It’s also good to clarify the actual limitations of each platform rather than accepting the marketing segmentation.

1 Like
2

What is your specific workload? Benchmark that workload and determine if 12 memory channels is more important vs 8 and higher clock speed.

3

There might be something strange going on specifically with the 9575F, in the CFD benchmark thread the 9575F got an unusually low score.

I’m not sure what is causing this issue but I’d suspect some kind of really unoptimized default NUMA execution happening that could be rectified with numactl.

4

I think the R23 single-thread score of 1,052 points is clearly wrong, because in the same benchmark the multi-thread score was 111,149 points. That’s 105 times higher than the single score, which is impossible for a 64-core system with SMT off.

Even so, the Blender benchmark results such as the Index of Nearest are quite puzzling. Regarding the differences between the EPYC 9575F and the TR 9985WX, aside from the well-known difference of 12-channel vs. 8-channel memory support, one major distinction, as Wendel pointed out in the linked source, is that the EPYC 9575F supports GMI-wide mode. This effectively doubles the per-CCD memory read bandwidth (writes are less certain). The TR 9985WX, despite having the same CCD layout, to my knowledge does not support GMI-wide mode. Still, I find it hard to believe that this would have a strong negative impact on performance.