AMD Ryzen 9 3900XT vs. Intel Core i9 10900K Linux Gaming Performance

[eva2000]

Thanks for sharing. Will be interesting to see how they compare over time as Kernel software mitigations accumulate over time and add some performance overhead

[lowflyer]

As I see it, this test run was less of a showdown between AMD and Intel. This test shows a lot better which game has managed to parallelize their code better.

[creative]

For gaming right now the vanilla 3600 is the sweet spot and great budget all arounder. Going to be a little while before I want to replace this thing. Pretty much am aiming for a 4700X after it comes out after a bit and there is an ok prices drop.

[vladpetric]

Gaming is Intel's last remaining strong point. Mostly for 2 reasons:

• It tends to take advantage of the higher single-core boost clocks
• It can be more memory latency sensitive than many other tasks

You can see the latter by comparing Renoir vs Matisse, where the non-chiplet version of the architecture has improved latencies and performs better even with massively reduced L3 caches.

Can you include a link for Renoir vs Matisse? Thanks.

[vladpetric]

Agreed, although differences that large in specific apps often means that they are taking different code paths.

Ok, there's a lot to be said here, but some of it is my speculation (in the absence of evidence).

* Typically most firms don't like doing code specialization, because it's one more dimension that they have to worry about when testing. Libraries do it (those libraries tend to be shared by multiple games). In any case, the specialization tends is done via the CPUID instruction, which primarily identifies available instruction sets (e.g., AVX2 present or not). Over there, the Ryzen line started with a pretty good feature set, but unfortunately parts of the implementation were pretty slow. It takes a lot of micro-ops and many cycles to execute some AVX2 instructions for instance (all the vector gather scatter instructions are super slow, also BMI2 pdep/pext; so if you activate a path that says - hey! AVX2 is there - you might actually get slower overall code ... ). See https://www.agner.org/optimize/instruction_tables.pdf - only covers the first generation Ryzen though ...

Why am I saying all this?

a) It could technically be the same code path, except that it doesn't run as fast on AMD.

b) Most importantly, if you have a more recent Ryzen processors (2000, 3000, or similar Threadrippers), I beg you to contact dr. Fog (on the forum at agner.org) and let him run his instructions benchmark suite on your processor. This would not only provide real evidence for this, but would help the larger community. He is not only good, but the kind of person who brings actual science to computer science

* Intel oftentimes sends people out to software firms to help them optimize their software on Intel CPUs. AMD doesn't do it. Typically those optimizations don't result in an overall slower code when run on AMD, though it does help with relative Intel CPU performance.

* For a lot of the benchmarks that are close to 0% (really, few people care about 2% change in games), it could be that the CPU is not getting taxed. Again, hypothesis/speculation.

[Michael]

It would be interesting to see a few results with the latest Proton version, I don't know which games would be best for benchmarking. Proton does have some extra threading work done on it so who knows, it might benefit AMD.

You could even run the proton version of linux port games to see if there is a difference in the way threading is done on Linux for some of these games.

This was the latest Proton 5.0 on Steam.

[Michael]

i don't see or missed it some where in the article, did michael say what speed memory was running on the test systems? Gaming performance on ryzen 3000 seems to benefit a lot from ddr4-3600 (or 3733/3800 if overclocking infinity fabric).

ddr4-3600

[smitty3268]

You're right, it doesn't, so scratch that.

Looking at the top of the chart shows some pretty significant differences between between Intel and AMD though.

Gaming is Intel's last remaining strong point. Mostly for 2 reasons:

• It tends to take advantage of the higher single-core boost clocks
• It can be more memory latency sensitive than many other tasks

You can see the latter by comparing Renoir vs Matisse, where the non-chiplet version of the architecture has improved latencies and performs better even with massively reduced L3 caches.

[bridgman]

Looking at the top of the chart shows some pretty significant differences between between Intel and AMD though.

Agreed, although differences that large in specific apps often means that they are taking different code paths.

[nranger]

I don't see or missed it some where in the article, did Michael say what speed memory was running on the test systems? Gaming performance on Ryzen 3000 seems to benefit a lot from DDR4-3600 (or 3733/3800 if overclocking infinity fabric).