Ryzen 7950X Optimization

[cardogab7341]

I recently upgraded my PC from an I7-3820 to a Ryzen 7950X. On the Intel machine, Prime 95 ran 4 workers, one on each core. When I installed Prime95 on the new machine, it is running 4 workers on 4 cores, with the rest of the cores as helpers. My question is, is this the optimum configuration? Has the 7950X been characterized to come up with the most efficient affinity profile? My work type is set to Let GIMPS decide.

Thanks

[Jurzal]

Originally Posted by cardogab7341

I recently upgraded my PC from an I7-3820 to a Ryzen 7950X. On the Intel machine, Prime 95 ran 4 workers, one on each core. When I installed Prime95 on the new machine, it is running 4 workers on 4 cores, with the rest of the cores as helpers. My question is, is this the optimum configuration? Has the 7950X been characterized to come up with the most efficient affinity profile? My work type is set to Let GIMPS decide.

Thanks

Congrats on your upgrade!

7950X is dual chiplet CPU with 32 MB of cache per 8 cores, so 2x8 cores with 2x32 MB of cache.
I have 5900X and have similar situation with 12 cores, 2x6, 2x32 MB of cache. So, optimally you want to have optimally 2 workers - one for first 8 cores, other one for other 8.

Currently my system runs 1 worker on DC LL, since it mostly is done within cache, second worker is set to PRP, since it takes most of the memory bandwidth. You can do P-1 if you want.

Running both workers on either P-1 or PRP not really effective, since RAM bottlenecks the system and CPU runs more on empty.

[Jurzal]

Originally Posted by Jurzal

Congrats on your upgrade!

7950X is dual chiplet CPU with 32 MB of cache per 8 cores, so 2x8 cores with 2x32 MB of cache.
I have 5900X and have similar situation with 12 cores, 2x6, 2x32 MB of cache. So, optimally you want to have optimally 2 workers - one for first 8 cores, other one for other 8.

Currently my system runs 1 worker on DC LL, since it mostly is done within cache, second worker is set to PRP, since it takes most of the memory bandwidth. You can do P-1 if you want.

Running both workers on either P-1 or PRP not really effective, since RAM bottlenecks the system and CPU runs more on empty.

Here is a screenshot of 5900X at work

[cardogab7341]

Originally Posted by Jurzal

Here is a screenshot of 5900X at work

Thanks for the tips I will investigate re-organizing work types. Can you show a screenshot showing the ms/iteration for the PRP test?

Thanks.

[Jurzal]

Originally Posted by cardogab7341

Thanks for the tips I will investigate re-organizing work types. Can you show a screenshot showing the ms/iteration for the PRP test?

Thanks.

No problem, here you go. If at idle, it usually rests at 3.6 - 3.7, if I have youtube open, discord or whatever, then around 3.9 - 4.0 for PRP. Yours should be much faster, because of DDR5 and faster/more cores. I tried out different setups - LL + LL, PRP + PRP, etc and I found LL + PRP most effective for my Zen 3 system.

My memory bandwidth typically maxes out at 25-28 Gbps, I am using 4x8 GB RAM at 3800 CL16 with tuned subtimings.

[cardogab7341]

I did some quick tweaking, and it looks like the optimum configuration for PRP in the 116m range for me, running a 6272k FFT is 2 workers and 14 helpers, split evenly between chiplets (1 worker, 7 helpers per chiplet). I get around 3.6ms/iteration on one and 3.9ms/iteration on the other.

Thanks for the tips.