New Ryzen CPUs faster than last gen GPUs?

[Jonazz]

After running an AMD 3970X (and some last gen GPUs) for a month I wanted to add my perspective.
It's a great CPU for me. I'm usually running many Docker containers (often within a VM) for my development work. After years of dual Xeon boards I made the switch to a single CPU because it has an ample amount of cores and I'm glad I did (increased performance, smaller board size, more mainstream, less pricey).
For F@H, I've tried 1 or 2 CPU slots with 32 cores each and total results are 500K to 700K PPD either way. I guess if I enabled some overclocking and fiddled with settings I might get it closer to 1M PPD. The CPU uses about 280W and hovers around 75 deg C (custom WC loop, liquid metal TIM). It drives my fans and pumps to full speed and I can definitely hear the system working. In contrast, my 3 GV100 cards combined use less than 400W and output 7M+ PPD at ~45 deg C (system stays whisper quiet).
So, from my experience I'll have to answer the topic with 'no'.
However, I'll keep running the CPU slots for a while longer because I really want to support moonshot with everything I got.

That's the spirit!

[Ichbin3]

my 3 GV100 cards combined

that's impressive ...

[maigner]

Currently, 32 is the max threads per slot configurable AFAIK. Setting anything higher in FAHControl reverts to -1 (which in turn translates to 29 for me). That's why I tried an additional slot with another 32 threads. Does anyone know if there's a way to set a higher thread count?

[Joe_H]

In WIndows, No. Limitations of Windows licensing, need to be compiled with correct parameters, and so on. Not an issue running under Linux, and possibly OS X. Though I don't know of anyone who has tested F@h on one of the 24c/48t or 28c/56t Mac Pros or the 18c/36t iMac Pro.

[MeeLee]

With 64 threads, you don't need an extra core for GPU.
Your system will run well crunching 64 threads at a time, plus GPUs.
It has sufficient overhead on each core to do 1 or 2 extra WUs (over-commitment), though I wouldn't put it any higher than 2x 32 threads I would say..

[mrpdaemon]

Just for reference, I'm folding with a Ryzen 3950x (stock speeds, no OC on CPU, RAM running at 3200Mhz) and I average 570k PPD, with lows around 480k and highs up to 710k. I'm running 30 threads since I also GPU fold on the same system, but 32 threads vs 30 threads seem to make a very marginal difference if any.

[MeeLee]

Just for reference, I'm folding with a Ryzen 3950x (stock speeds, no OC on CPU, RAM running at 3200Mhz) and I average 570k PPD, with lows around 480k and highs up to 710k. I'm running 30 threads since I also GPU fold on the same system, but 32 threads vs 30 threads seem to make a very marginal difference if any.

Thanks!
Seems like the same thing like on my Xeon.
18 threads vs 20 threads was almost the same.
Also task manager showed 100% CPU utilization on either one of the settings (this was without GPU).
I would guess in some way that FAH over commits the CPU by a little. Perhaps due to 1 CPU thread being utilized to feed the other threads (much like it would feeding a GPU).

[mrpdaemon]

[
Thanks!
Seems like the same thing like on my Xeon.
18 threads vs 20 threads was almost the same.
Also task manager showed 100% CPU utilization on either one of the settings (this was without GPU).
I would guess in some way that FAH over commits the CPU by a little. Perhaps due to 1 CPU thread being utilized to feed the other threads (much like it would feeding a GPU).

Most CPU's are able to boost to higher clock speeds when using less threads, so for example going from N cores to 2N cores will not give you exactly 2X linear performance increase. Ryzen series are particularly effective at boosting, so you do get similar levels of performance especially at higher thread counts (28 vs. 30 vs 32 makes no difference in my experience). But I noticed using more threads does give power efficiency improvements (lower clock speeds, slightly less power draw) and lower temps.

[NBR]

I wonder how next-gen GPUs will fare.