Summing Up AMD Ryzen 9 3900XT vs. Core i9 10900K, Ryzen 5 3600XT vs. Core i5 10600K

[phoronix]

Phoronix: Summing Up AMD Ryzen 9 3900XT vs. Core i9 10900K, Ryzen 5 3600XT vs. Core i5 10600K

Complementing this morning's AMD Ryzen 5 3600XT / Ryzen 7 3800XT / Ryzen 9 3900XT Linux benchmarks, here is a side-by-side look at the Ryzen 9 3900XT up against the Core i9 10900K and the Ryzen 5 3600XT up against the Core i5 10600K for these competing processors. This is a quick look at how these competing models stack up in the 130+ benchmarks utilized so far...

http://www.phoronix.com/scan.php?page=news_item&px=Side-By-Side-XT-Comet-Lake

[Beherit]

Thx Michael. Much appreciated that you write up a summary.

[morydris]

Yes, this is a fantastic way to present the [bewildering] results. Results are really close these days that being able to zoom in on applications closer to my workload is helpful (unfortunately I don't think Place&Route is modeled well by any of these).

[Michael]

Yes, this is a fantastic way to present the [bewildering] results. Results are really close these days that being able to zoom in on applications closer to my workload is helpful (unfortunately I don't think Place&Route is modeled well by any of these).

Are there any place and route type benchmarks out there?

[morydris]

I'm not aware of anything ... yet ... but the Yosys + nextpnr-ecp5 packages are quite portable so perhaps something could be produced. I'll let you know if this materializes.

After my comment I looked at the list again and there is one concern: there really are three classes of benchmarks here.

Kernel-style application where the performance a medium/small loop is all that matters - this is almost synthetic. Some of these show a dramatically difference, but are highly sensitive to how well they were tuned for a particular processor.

Secondly, highly multi-threaded application will almost certainly favor Ryzen.

Finally, there are the common class of lightly threaded applications.

I know this is a lot to ask, but it really would be more informative if we could clearly distinguish these. For the latter two, I'd suggested if possible measuring the geometric mean of the load average and classify everything <= 2 in the "single thread" category. For the first category you really just have to know. Eg., CoreMark, compression, FFT, Nettle, HMMer, DGEMM, Gimp unsharpen mask, ... are clearly kernels. I do _not_ consider codecs (like AV1 and H.264) to be kernels.

UPDATE: I missed that this was a summary of a much larger story. Sorry. My comment still holds though, but the original story has a lot more detail.