Linux Performance Benchmarks for 13900K
www.phoronix.com
The 13900K just don't play nice on Linux
View attachment 69890
Geometric Mean
View attachment 69891
Sifting through the article, it looks like a few things bubble to the surface:
AMD's decision to implement AVX-512 in Zen4 is paying dividends where it is supported. In heavy AVX-512 benchmarks, ZEN4 gains a lot of performance over Raptor Lake. This isn't a case where Intel could have changed the outcome, no matter of their approach. If they left AVX-512 enabled, but restricted threads to the P cores, they would be down many cores to the higher end Zen4 parts and would look bad against even the 7700x (and actually loose to it a couple of times in heavy MT benches while using alternate code paths). If they added AVX-512 to the E cores to help those benches, it would balloon those cores and the math on die space wouldn't work. If they decided to just go with 12 P-cores, it might have come out ahead in a couple of tests where AVX-512 is relevant, assuming that there are specific conditions in place, but, they would still be down by 4 cores vs. the 7950x and take a lot of losses.
There are multiple benches where the 7900 and 7950x loose significant performance against the 7700x and 7600x. There's a penalty that's still present in some benches for having multiple CCDs. It's not drastic overall, but it is notable. There are some tests where the 7700x and 7600x beat the 13900K and the 7950x and 7900x loose to it. That could likely be addresses through affinity adjustments, thought that's just speculation.
Overall power usage, while achieving higher performance, across a WIDE variety of tests, favors ZEN4, HOWEVER, this is not a huge difference as the 7950x is chewing through a lot of power itself to get the wins that it gets, just not usually as much as Raptor Lake.
Linux, especially the most recent kernels, seems to do a solid job of managing the E cores. In some cases, it seems to do better managing the E cores than it does handling the multiple CCDs of the two CCD Zen4 parts.
I speculate that, if AMD does release dual CCD X3D parts, and even with the single CCD parts, Raptor lake will suffer heavily in many of these benchmarks. I make this statement because I note that the 5800X3d is often nipping at the heels of the 13900K and even beats it on a few benchmarks. This is while the 5800X3d is suffering a massive penalty in single and all core boost on top of having a smaller L2, no AVX-512 support, and not having the other Zen4 core improvements. The Geomean for the 5800X3d is barely over the 5800X largely due to the clock deficit. From the information that we have, it looks like Zen4 will do much better in that regard.
All of this is to point out that the often speculated 13900KS release will likely need a significant clock speed bump to keep from being significantly outpaced, at least in the Phoronix Linux test suite.
Facebook
Twitter