CPU requirement for GPU folding (NV)

[foldinghomealone]

In our team we constantly discuss CPU requirements for NV GPU folding and the necessity for CPU-OC to 'feed' a NV GPU.
Mainly because when someone is setting up a new rig he wants to lower costs and use the cheapest CPU.

I was reading in several posts here that CPU speed for GPU folding is not important because for one GPU client one CPU thread would be enough because it is 'just' used to keep the I/O channel between CPU and GPU open. There would be no difference if the thread runs at 2GHz or 5GHz. (may be my own words)

My own experience goes in this direction as I tried to down-clock and OC my CPU in several different GPU projects and didn't see much difference.
But this opinion is not shared at all in our team. Most of our members are the opinion that the CPU can slow down GPU folding significantly when there is not enough CPU horesepower.

Maybe you can give some insights:
- Does CPU matter? Can a current non-OC celeron utilize a single Titan-X (P) (crazy example, I know) or do I need an OCed i7?
- If it matters are there pairing recommendations?
- When I have one GPU slot I see that the folding process 'jumps' from one thread to another. Does this 'jumping' cost folding power? Should I lock the folding process to one certain thread?
- Does the turbo-clocking need some time until it turbo-clocks and is therefore losing folding power? (This answer could be interesting for CPU-folding as well.)

Many questions, i hope you can understand them. Many thanks in advance

[ComputerGenie]

I can't help much, other than to say that there's no detectable difference between a GTX 1080 set up on a board that is running a AMD FX-8320 @ 3.25 GHz and a GTX 1080 set up on a board that is running a i7-6700K @ 4.5GHz

[boristsybin]

You need at least one cpu thread for one GPU; but the best practice is one cpu core for one GPU.
So, i think 3-y.o. dual-threaded celeron is enough for even titan x pascal, but only single titan.
AMD GPUs are not so CPU-dependable as NV (as for my experience with R9 290x, RX480 and GTX1070).

i`m succesfully use dual 1070 with dual-core (four-threaded) i3 4170 running round o`clock; CPU load fills up to 100%, 50% of is for FahCore_*, but i feel no discomfort even while routine work on my system (browsing, msoffice, whatever...) and with kaspersky AV

[rwh202]

Yep, as long as each slot has got a core, it doesn't really matter how slow it is. I tend to use pentiums (2C 2T) with dual GPUs (up to 1080s so far) and don't think I'm losing anything compared to my machines with i7s.
On a slower GPU (1060) with windows, it was easy enough to drop the processor speed by 50% using the power management settings and there was no discernible increase in TPF.

[foldy]

The only thing to check is the mainboard pcie slot speeds, you don't want to go below pcie 3.0 x4 on fast GPUs and pcie 3.0 x8 or pcie 2.0 x16 has max performance on Windows. For slow GPUs or Linux that doesn't matter.

[JT3rd]

As a newbie, I started off with an FX-8350, both undervolted and underclocked.

Using this CPU with the GTX 1050 & TI, running the CPU from 2.0 to 4.4GH there were measurable GPU performance increases with the higher clock speeds. If I assigned a GPU to a single core, it would use the entire core until I hit about 3.2GH, at this point there was a little bit of overhead to spare. From my perspective (using Heaven, GPU-Test, Cinebench, FAH-Bench) at the lower clocks the GPUs were bottlenecked by the CPU, and because this CPU is such a power hog, I opted to get a Pentium G4560, with higher single core performance.

With the new G4560, running 2 GPUs on Windows 7, CPU usage hovers at 50% and the system is still snappy

Having said all this, regardless of the lower clock speeds, while the bench-marking tools were definitely degraded, the performance of the actual FAH client did not appear to be as degraded as the benchmarks.

[road-runner]

i3-4170 Haswell Dual-Core and GA-Z97X-Gaming 3 LGA 1150 Intel Z97 HDMI SATA 6Gb/s USB 3.0 ATX Intel Motherboard runs 2-1080s just fine...

[Nathan_P]

Xeon E3-1230L v3 @ 1.8ghz runs my 1080 and 1070 just fine and the bonus is its only a 25w tdp cpu, the rig is usually outputting 1.4-1.5m PPD

[ChristianVirtual]

@OP

You could see at ppd.fahmm.net for project 9414: a pentium G2020 at 2.9 GHz "outperform" a i7-6700 at 3.4Ghz, both with a 1080 and Linux . Unfortunate i don't know what the OC setting for both 1080 are (as not reported in log files) but I think it's fair to conclude that the CPU speed has less impact on the performance compared to the GPU itself.

[v00d00]

For me personally, running GPU on Linux. I run each GPU instance separately. One fahclient per card. Each fahclient is bound to one cpu core. Those cores are isolated at kernel level so they are no longer bound to the scheduler (so nothing else will use those cores except FAH), this is done by passing isolcpus=x via the grub or lilo bootloader at startup, then using taskset to bind the clients to the isolated cores when my start scripts start everything up.

I use an AMD FX 8320e @ stock clocks, run currently a GTX 1050 Ti (180k PPD), and a 6 core SMP. Slowdown was noticed when the GPU instance(s) werent placed on isolated cores, presumably due to the fight between SMP and GPU for cpu time. PPD was increased after I fixed it.

X always uses about 10-15% of core 0 when GPU is running, even if you use a different card for display tasks. But if you run the box from text mode it still folds without issue. Havent worked out why X is using so much time when my folding card(s) arent even used for any video rendering.

For the sake of completeness. It isnt possible to isolate cores for SMP using isolcpus, as far as im aware, since it ends up binding to a single core and spawns threads which are also bound to that core, even when taskset across multiple cores. But you can taskset an SMP instance to multiple cores using taskset if the cores are still handled by the scheduler.

[foldy]

On GPUGRID they have an option for nvidia GPUs to use the poll approach or the wait approach. The default poll approach results in higher GPU speed.
There is currently an ongoing discussion about that on OpenMM. https://github.com/pandegroup/openmm/issues/1541

[Ricorocks]

Foldy wrote,

The only thing to check is the mainboard pcie slot speeds, you don't want to go below pcie 3.0 x4 on fast GPUs and pcie 3.0 x8 or pcie 2.0 x16 has max performance on Windows. For slow GPUs or Linux that doesn't matter.

I'm in the hunt for a 'windows' machine, that will have duel NV 1070s. So PCIE 2.0 x 16 slot is fine with 2 1070's? I'll search some more using cpu i3-4170

Thanks
Rick

[foldy]

Looks good.

[Ricorocks]

Looks like allot of good choices for CPU's, by picking more efficient from the list.

https://www.cpubenchmark.net/power_performance.html#id2

[foldy]

If you also want to fold on CPU then be sure it supports AVX for highest PPD. And each GPU needs one CPU core/thread to feed it.