Folding Forum

Hi,

I have just ordered a new system for my work pc, I use solidworks and the cfd solver can use all the cores. My question is how should I set up FAH, how many cpu slots of how many cores? I have also ordered 2 2060 supers to sit in the build just for folding.

Automatic selection of FAH slots should be fine.

My 3990X works only with 32 threads..
i put the second CPU slot.. and become 2x 16 Threads.
actually the power consumption with 32 threads are so so same like the consumption in full load with 128 threads... (WTF!!!)
other bug.. i have RTX 4000 - but the system cant match this card...

if have some updated client (this one is from 2018...) will be better.
cheers

On my "little" threadripper, I've found that 1 slot per "chiplet" has worked best (better?), then use Process Lasso to keep each FAH core, so each core is kept to its own L3 cache (this is the important part).

With GPUs in the system, you want to leave a couple CPUs somewhere "available" to tend do the GPU overhead too.
I've setup 4 slots on my PC:
2 CPU slots with "10" CPUs each. (I have not noticed PPD differences with values between 10 and 16 here, there are only 8 "real" processors per chiplet anyway)
2 GPU slots auto configured (for a 2080TI and a 2070 super).

I've then installed Process Lasso (a VERY useful tool for these ridiculous processors), "right clicked" on one of each "type" of process name for FAH (currently fahcore_22 and fahcore a7) and used "Instance Balancer" to create "Equal CPUs per instance" for each.
I know this can all be done on linux too, I just haven't done it yet so I can't speak to it.

The end result is 2 CPU slots that reasonably fully saturate the actual capability of each Chiplet (with some extra thread capacity around to make the GPUs happy), and 2 GPU slots that crank away without contending for resources.

If it was me (and it isn't, my budget tends to PCs under $100) I would set up 7 CPU slots of 16, and a CPU slot of 12.

While 3 slots of 32 cpus, 1 slot of 16, and a CPU slot of 12 would make more points in theory, I hear rumors that few Work Units are compatible with larger slots, so you might make more points and do less science.

Many WUs have issues with large primes and their multiples 7 is always large, 5 is sometimes large, and 3 never is.

so 32, 16 and 12 are all multiple of just 2 and 3.

so ive just got this up and running, but only 32 cores are recognised as mentioned above? any ideas?

Perhaps tha is a Quadro FX 4000 or something else. Please visit viewforum.php?f=83 and provide the necessary information.

This is the AMD Threadripper cpu 3990x 64 core 128 thread.

I read somewhere that windows is limited to 32 threads per slot, You should be able to add a second slot of 32 threads and so on. Linux would handle this a lot better as I think the limit is either 96 or 128 threads but I get that its your work machine and you need windows and no one is going to drop 4k on cpu just for folding.

If I add another slot then it gets split into 2 containing the same 32 in total.

scerbera wrote:If I add another slot then it gets split into 2 containing the same 32 in total.

hmm, thats odd. In theory you should be able to cpu slots upto the quantity of cpu cores. I don't have any hardware to test this on as my threadripper is only a baby 1920x. Someone else will have to chime in on this

As I recall, you need to use Configure in FAHControl and manually set the CPU thread count for the slots if you are setting up more than one. The default setting for CPU slots is '-1' which leaves the setting up to the software, and in this case it sees 32 total that it will apportion over more than one slot.

Have you considered changing the BIOS setting to disable sharing a single FPU on a pair of logical CPUs? (Intel "HyperThreading" or the AMD equivalent)? FAH throughput depends mostly on the number GFLOPS, not the number of threads.

I have turned smt or hyper threading off, so just 64 physical cores, but it still only recognises 32, so it runs with 29/28 cores now on a cpu project. The chip is still pushing out the tdp of 285 watts with no overclock, but it could be doing more. The tpf is running at 9 seconds roughly.

I finally get a full workload and run into a NUMA bottleneck.