Threadripper 9985WX cache and memory benchmark

[fizzgabiz]

While trying to optimise my memory bandwidth, I noticed that AIDA64 fires up 128 threads on a 9985WX and thus takes an eternity to finish the benchmark. This approach might have some sense regarding the cache performance, but it renders the test pretty useless for determining the memory module performance on an 8 channel interface. But I would actually prefer to stick to AIDA64 for comparable results from quite some years now. Is there any tweak available for this? I was able to get some meaningful results by shutting down 6 out of 8 cores on every CCD in the BIOS, but this is rather cumbersome and leaves me with the impression of an artificial result.

[Fiery]

17 minutes ago, fizzgabiz said:

While trying to optimise my memory bandwidth, I noticed that AIDA64 fires up 128 threads on a 9985WX and thus takes an eternity to finish the benchmark. This approach might have some sense regarding the cache performance, but it renders the test pretty useless for determining the memory module performance on an 8 channel interface. But I would actually prefer to stick to AIDA64 for comparable results from quite some years now. Is there any tweak available for this? I was able to get some meaningful results by shutting down 6 out of 8 cores on every CCD in the BIOS, but this is rather cumbersome and leaves me with the impression of an artificial result.

Make sure to upgrade to the latest beta version of AIDA64 Extreme available at:

https://www.aida64.com/downloads/latesta64xebeta

After upgrading to this new version, make sure to restart Windows to finalize the upgrade.

Let me know if it helps.

[fizzgabiz]

THX, Fiery, this acutally works "as usual", but the timings do not seem plausible and are not in line with other measures (latency >102ns for a current tune of RDIMM DDR5 6400-44-46-46-103). Should be ~65ns.

 

Memory Clocks seem generally a bit off and Motherboard->SPD is still empty in my case.

[Fiery]

23 hours ago, fizzgabiz said:

THX, Fiery, this acutally works "as usual", but the timings do not seem plausible and are not in line with other measures (latency >102ns for a current tune of RDIMM DDR5 6400-44-46-46-103). Should be ~65ns.

 

Memory Clocks seem generally a bit off and Motherboard->SPD is still empty in my case.

We've checked and memory latency is indeed around 100ns on similar configurations to yours.

[fizzgabiz]

Thank you, Fiery. Can you provide some insights in your algorithm? What is the probability to hit in an open row buffer? 56%?

[Fiery]

13 hours ago, fizzgabiz said:

Thank you, Fiery. Can you provide some insights in your algorithm? What is the probability to hit in an open row buffer? 56%?

We only provide deep insights into our algorithm under NDA.  What I can tell you is that we use a block-random algorithm that both AMD and Intel approves as a proper method to measure memory latency on their platforms.

[fizzgabiz]

OK, THX! I just tried to figure out, how this relates to other benchmark results. But at least it works now and I have a basis for comparison.