Does folding@home need anything extra configured to work well on AMD CPUs like MatLab does. I noticed even though I put the folding power for FAH at high only 3 arguably 4 cores get used. My GPU sees a lot of utilization at 46% but my CPU at 10% leading me to think it is being under utilized.
I am running a AMD Ryzen 7 3800x CPU. The GPU is a Nvidia GTX 980 Ti. In case it matters I have 64GB memory and I have FAH installed and data set to a 10k rpm hdd. Not really sure what kind of storage behavior FAH has and didn't want to thrash any of my SSDs.
The issue MatLab has with AMD CPU I am talking about and how to fix it is detailed in this blog. (Apparently I am too new to post links. Search AMD Matlab performance and it should be too hard to find.)
Performance on modern AMD CPUs
Moderator: Site Moderators
Forum rules
Please read the forum rules before posting.
Please read the forum rules before posting.
-
davidcoton
- Posts: 1094
- Joined: Wed Nov 05, 2008 3:19 pm
- Location: Cambridge, UK
Re: Performance on modern AMD CPUs
Are you quoting the peak load, or the overall figure? There are server overload issues at present which make it hard to keep slots in work.
The default client configuration is usually pretty much optimal. Experience has shown that one CPU core is needed to support each GPU slot, and overall throughput declines if the CPU is fully allocated to a CPU WU. The GPU's utilization varies between projects, more than 50% should be normal. I would expect CPU utilization to peak at 100% for the committed cores (ie 75% overall for 3C on a 4C CPU). It may be lower for a first WU, there is a minor issue that results in a first CPU WU often being for a single core only.
The default client configuration is usually pretty much optimal. Experience has shown that one CPU core is needed to support each GPU slot, and overall throughput declines if the CPU is fully allocated to a CPU WU. The GPU's utilization varies between projects, more than 50% should be normal. I would expect CPU utilization to peak at 100% for the committed cores (ie 75% overall for 3C on a 4C CPU). It may be lower for a first WU, there is a minor issue that results in a first CPU WU often being for a single core only.
-
JimboPalmer
- Posts: 2522
- Joined: Mon Feb 16, 2009 4:12 am
- Location: Greenwood MS USA
Re: Performance on modern AMD CPUs
Welcome to Folding@Home!
The software does a pretty good job of optimizing.
For CPUs, Core_a7 is actually two code bases, for older CPUs it uses SSE2, for newer CPUs is uses AVX_256. It declares which it is using as it starts each Work Unit. All the Ryzen CPUs should use the latest, avx_256. (as well as some older AMD CPUs)
For Intel chips it is about the First generation of Core i3/5/7 that it switches to avx_256.
https://en.wikipedia.org/wiki/SSE2
https://en.wikipedia.org/wiki/Advanced_ ... Extensions
For GPU utilization I thing GPU-Z's sensor screen is very good.
https://www.techpowerup.com/gpuz/
The latest AMD GPUs require a Core_22 Work Unit, before Navi or RDNA, both AMD and Nvidia GPUs could use Core_21. F@H is transitioning to Core_22 over time. (newer code does better science, but existing projects will finish on Core_21)
F@H has no real storage bottlenecks, it writes a checkpoint every so often, but does not wait on it.
Nvidia seems to do a busy wait to feed the PCIE bus for GPU AMD uses interrupts so wastes less CPU time.
RDNAs OpenCL implementation caused a lot of issues across applications.
The software does a pretty good job of optimizing.
For CPUs, Core_a7 is actually two code bases, for older CPUs it uses SSE2, for newer CPUs is uses AVX_256. It declares which it is using as it starts each Work Unit. All the Ryzen CPUs should use the latest, avx_256. (as well as some older AMD CPUs)
For Intel chips it is about the First generation of Core i3/5/7 that it switches to avx_256.
https://en.wikipedia.org/wiki/SSE2
https://en.wikipedia.org/wiki/Advanced_ ... Extensions
For GPU utilization I thing GPU-Z's sensor screen is very good.
https://www.techpowerup.com/gpuz/
The latest AMD GPUs require a Core_22 Work Unit, before Navi or RDNA, both AMD and Nvidia GPUs could use Core_21. F@H is transitioning to Core_22 over time. (newer code does better science, but existing projects will finish on Core_21)
F@H has no real storage bottlenecks, it writes a checkpoint every so often, but does not wait on it.
Nvidia seems to do a busy wait to feed the PCIE bus for GPU AMD uses interrupts so wastes less CPU time.
RDNAs OpenCL implementation caused a lot of issues across applications.
Tsar of all the Rushers
I tried to remain childlike, all I achieved was childish.
A friend to those who want no friends
I tried to remain childlike, all I achieved was childish.
A friend to those who want no friends
Re: Performance on modern AMD CPUs
Good advices above.
It sounds like your CPU did not have an active CPU task running. Check in FAHControl.
I'm running Linux Mint on my 3950X and give 30 threads to FAHClient. I see 100% utilization across all 30 threads and it folds like crazy. The last two CPU threads are for GPU tasks and for still being able to use the computer while folding. Zen 2 is amazing; I can hardly tell while using my computer that 30 of my 32 threads are floored!
I have CPU folding on 24/7 but turn off GPU folding if I want to use my computer for getting work done.
It sounds like your CPU did not have an active CPU task running. Check in FAHControl.
I'm running Linux Mint on my 3950X and give 30 threads to FAHClient. I see 100% utilization across all 30 threads and it folds like crazy. The last two CPU threads are for GPU tasks and for still being able to use the computer while folding. Zen 2 is amazing; I can hardly tell while using my computer that 30 of my 32 threads are floored!
I have CPU folding on 24/7 but turn off GPU folding if I want to use my computer for getting work done.
Re: Performance on modern AMD CPUs
I have just started folding and the PPD figures initially didn't encourage to use CPU; until I double-checked my new 3950X with 30 threads.
Estimated PPD: 433k (consuming 186w)
For comparison:
GTX 970 - 353k (consuming 135w)
GTX 1070 - 742k (consuming 150w)
GTX 2070 - (1.3M) (consuming 175w)
So its worse per watt, but not order by magnitude worse. Bit below 1/3 of GTX 2070 with equal watts.
Estimated PPD: 433k (consuming 186w)
For comparison:
GTX 970 - 353k (consuming 135w)
GTX 1070 - 742k (consuming 150w)
GTX 2070 - (1.3M) (consuming 175w)
So its worse per watt, but not order by magnitude worse. Bit below 1/3 of GTX 2070 with equal watts.
Re: Performance on modern AMD CPUs
Yes, that is the nature of the beast when comparing CPUs with GPUs in regards to floating point operations. GPUs by nature of their design are optimized for these types of calculations.
For F@H, GPUs are the winner but with projects such as rosetta@home, the 3950X is an attainable king. Due to the complexity of creating the work, R@H only issues work units for CPUs.
The thing I like about the 3950X is that even though I have 30 cores running at 100%, I don't even notice it when using the computer. A phenomenal processor! The GPU? I really notice when that thing is processing.
For F@H, GPUs are the winner but with projects such as rosetta@home, the 3950X is an attainable king. Due to the complexity of creating the work, R@H only issues work units for CPUs.
The thing I like about the 3950X is that even though I have 30 cores running at 100%, I don't even notice it when using the computer. A phenomenal processor! The GPU? I really notice when that thing is processing.