FYI: No significant change with 24H2 vs 23H2 for Zen 3's R9 5900X in CPU Profile or Steel Nomad

[Migz - DH]

Greetings,

As you may know, AMD and Microsoft have finally made significant progress on Zen CPU scheduling in Windows 11. I rejoined the Windows Insider program to try out the difference in 24H2.

Unfortunately from a testing perspective, I found no difference in 3DMark's CPU Profile and Steel Nomad on my Zen 3 Ryzen 9 5900X.

I am not saying that anything is wrong with 3DMark's benchmarks. They just don't seem to be useful for finding the performance difference for this OS change. * Perhaps older 3DMark benchmarks would show a difference.

Cheers


* I am not surprised that CPU Profile did not show a difference, because that test maxes out threads. I suppose the reason Steel Nomad does not show a difference is that it is designed to stress the GPU, not the CPU.

[Migz - DH]

Regular Fire Strike shows a significant difference since the last time that I ran it, but that was nearly 3 years ago, and I have made a lot of BIOS changes since then. Still, maybe someone else can run regular Fire Strike before upgrading to 24H2 as well as afterwards and compare the results.

[surgicalCFe]

With only the power limit for my 40 series gpu maxed (110% in afterburner) my Steel Nomad score before 24h2 was sub 9000 (~8800) and after 24h2 I'm at 9200. I'd say that's a substantial uplift. I haven't give SN an OC'd run yet since 24h2 but before 24h2 my best was ~9800. I'll do some OC'd runs eventually, I tinker not super into it but I'm betting I'll break 10k.

[Migz - DH]

Publicado originalmente por surgicalCFe:

With only the power limit for my 40 series gpu maxed (110% in afterburner) my Steel Nomad score before 24h2 was sub 9000 (~8800) and after 24h2 I'm at 9200. I'd say that's a substantial uplift. I haven't give SN an OC'd run yet since 24h2 but before 24h2 my best was ~9800. I'll do some OC'd runs eventually, I tinker not super into it but I'm betting I'll break 10k.

Very interesting. Thanks for your comment.
What CPU are you using? Is that with a regular 4090?

[UL_Jarnis]

Older tests like Fire Strike and non-extreme Time Spy do not thread to max cores on high core count systems (in general they do not thread much beyond 8 cores). That can cause loss of performance on multi-CCD AMD chips and more-than-8-core Intel chips. As we understand this, the Windows change mentioned will try to keep the load in these scenarios to not be trashing between CCDs and cores in a way that would degrade performance, but this is just an educated guess - we have no real visibility to the internals of Windows kernel in issues like this.

This is just benchmark behaving exactly like older games would behave with a simplified threading model that, when core counts increase, gets swamped by the overhead.

Time Spy Extreme CPU test and CPU Profile test are more recent and are not affected by this. Port Royal, Speed Way and Steel Nomad GPU tests are also unaffected.

Could we "fix" older tests? Sure, but that would invalidate all old results and in a way having an older test to behave exactly like older games would do is useful, and we generally do not want to do workload changes that invalidate all old data. Depending on what you want, just use the correct test. 3DMark has many to choose from exactly for reasons like this.

[Migz - DH]

Publicado originalmente por UL_Jarnis:

Could we "fix" older tests? Sure, but that would invalidate all old results and in a way having an older test to behave exactly like older games would do is useful, and we generally do not want to do workload changes that invalidate all old data. Depending on what you want, just use the correct test. 3DMark has many to choose from exactly for reasons like this.

I absolutely do not suggest changing the tests. :-)