PPD Database - HFM exports needed

[skydivingcatfan]

It is still kind of rough.

[foldinghomealone]

New data for:
GTX 980
i7-3770k @6t
R5 3600
RTX 2070 Super
RTX 2080 Ti
RX 580
TR 2950x @12t

Thanks to:
skydivingcatfan
lafrad

[lafrad]

Swapped my threadripper over to a single big slot and see if that changes the aggregate PPD. Not anticipating a big difference but who knows?!?

[lafrad]

Oyi, CPU slots to 30 with the 2 GPU's caused *way* too much contention, lowered PPD and slowed the system considerably.

Back to a single 24 thread CPU slot along with the 2GPU's. That's leaving "6 threads" idle, but I have a feeling this will be better than the 30 slot core.

[PantherX]

Oyi, CPU slots to 30 with the 2 GPU's caused *way* too much contention, lowered PPD and slowed the system considerably...

Weird... on a 32 CPU system, the ideal settings would be:
2 CPUs for 2 GPUs (Nvidia)
30 CPUs

If you have other processes running in the background, it would have negative impact on the system. Also, you could use affinity to lock the processes to particular CPUs for maximum efficiency.

You can also try 27 as it is another safe number: viewtopic.php?f=72&t=34350&start=45

[foldinghomealone]

Oyi, CPU slots to 30 with the 2 GPU's caused *way* too much contention, lowered PPD and slowed the system considerably.

Back to a single 24 thread CPU slot along with the 2GPU's. That's leaving "6 threads" idle, but I have a feeling this will be better than the 30 slot core.

Without comparing either a big sample size for average or WUs of the same project it's not really possible to tell.

That the system is slowed down when all threads are occupied... To be expected

[foldinghomealone]

New uploads for:
i7-7700
Quadro M2000
R5 3500U
R9 3900X
RTX 2070 Super
RX 580
RX Vega 64
RX Vega 8

Thanks to:
lafrad
AOD_N3URAL
absolut_zero3

[lafrad]

this is my main system, and there are other tasks that run around even when idle.
30 CPU threads and a single one for each GPU meant any other tasks would be fighting all of those FAH threads for CPU time, getting in their way. The amount of contention that caused for the GPU threads means they *each* dropped PPD more than the entire CPU could produce in the first place... (the 2080 Ti dropped almost 500K PPD for a few WUs)

I'll give 27 a shot to see if that ups the PPD of the CPU without impacting the GPUs. As it is, it still looks like 2 12-thread CPU slots is more PPD (and WUs) per day than anything "large".

If there was an easier way to do affinity locking on linux, I"m all ears, but when ALL the threads are used, it will always be a contention issue.

[MeeLee]

Not sure if a threadripper has the required CPU speed to feed an RTX 2080Ti.
It takes 3 to 3,5Ghz to keep a 2080Ti fed.
Not sure if FAH will utilize 2 threads when 1 thread doesn't supply the necessary data to feed the GPU.

[NoMoreQuarantine]

Submitted some 2060 KO data. It's looking pretty good! I wish I had a regular 2060 to compare it against as it has the same shading unit count, but is actually a binned TU104, not the normal TU106.

@skydivingcatfan very interesting chart! Not the results I was expecting; I thought the 2070 Super would be in the lead the first couple years, but it looks likes the 2060 Super is the clear winner.

[Joe_H]

Not sure if a threadripper has the required CPU speed to feed an RTX 2080Ti.
It takes 3 to 3,5Ghz to keep a 2080Ti fed.
Not sure if FAH will utilize 2 threads when 1 thread doesn't supply the necessary data to feed the GPU.

Please look up the MHz Myth, now more commonly known as the GHz Myth. Clock speed of a CPU only has an indirect effect on data throughput, more important is what generation and architecture the processor and its I/O channels.

[lafrad]

This threadripper is water-cooled and running 4.1 ghz all-core... should be good no matter what mythos you follow...

[skydivingcatfan]

@skydivingcatfan very interesting chart! Not the results I was expecting; I thought the 2070 Super would be in the lead the first couple years, but it looks likes the 2060 Super is the clear winner.

I would say there is nothing clear about the 2060 Super. If you look at the HFM database there is only a sample size of 1.

[MeeLee]

Not sure if a threadripper has the required CPU speed to feed an RTX 2080Ti.
It takes 3 to 3,5Ghz to keep a 2080Ti fed.
Not sure if FAH will utilize 2 threads when 1 thread doesn't supply the necessary data to feed the GPU.

Please look up the MHz Myth, now more commonly known as the GHz Myth. Clock speed of a CPU only has an indirect effect on data throughput, more important is what generation and architecture the processor and its I/O channels.

It's been tested on modern Intel and AMD CPUs a few times on this forum.
3,5Ghz is really the minimum for high end GPUs.
You might need the same speed for older CPUs pushing slower GPUs (like 2060/2070).
Linux HTOP will show clearly how much of the CPU is being used, and how much is idle data sent to the GPU.


In theory, threadrippers and Ryzens run 4,1Ghz. But they hardly ever do.
Only in tuned, short burst tests. Most motherboard VRMs don't support the power a threadripper needs to run at those frequencies.
Blame it on AMD marketing.
So yes, a threadripper can do 4,2Ghz at nearly 400-500Watts.
If you run it at the stock 280W setting, it will more than likely run in a 3,6-3,8Ghz ratio.
The higher core versions even less.
And if your cooling solution isn't powerful enough, it'll run even slower.
Even at 280W, a triple 360 water cooling solution will have trouble keeping up.
It's not unheard of, the top end Threadrippers running at only 2,9Ghz with a continuous load on all cores.

Best thing you can do, aside from buying a fat and decent cooling solution, is to make sure also the VRMs get extra cooling (a small 120mm case fan blowing on the heat sinks, should be enough). Because they tend to get burning hot!

[Joe_H]

It's been tested on modern Intel and AMD CPUs a few times on this forum.
3,5Ghz is really the minimum for high end GPUs.
You might need the same speed for older CPUs pushing slower GPUs (like 2060/2070).
Linux HTOP will show clearly how much of the CPU is being used, and how much is idle data sent to the GPU.

Exactly where? What has been tested is that some low end CPUs running at sub-2 GHz speeds did result in reduced throughput on high end GPUs. That does not equal showing that some magical minimum GHZ is needed for a GPU.