Run two tasks per GPU?

Moderators: Site Moderators, FAHC Science Team

Post Reply
Lamberto Vitali
Posts: 83
Joined: Fri Feb 25, 2022 12:21 am

Run two tasks per GPU?

Post by Lamberto Vitali »

In Boinc, if I have a GPU not maxed out (because the CPU isn't powerful enough for example), I can run 2 tasks on the GPU at once, then each task is using a different CPU core to help out. Not only does that give me twice the cores to help the GPU, but it means while one task is stalled waiting for the CPU, the other can use the whole GPU. I end up with 100% GPU utilization.

How do I do this on Folding? I've checked the list produced by typing "FAHClient --help" but I can't find anything in there.
PaulTV
Posts: 179
Joined: Mon Jan 25, 2021 4:53 pm
Location: Netherlands

Re: Run two tasks per GPU?

Post by PaulTV »

FAH can only run one job per GPU. I guess you can have 1 Boinc job and 1 FAH job running concurrently.
Image

Ryzen 5800X / RTX 4090 / Windows 11
Ryzen 5600X / RTX 3070 Ti / Ubuntu 20.04
Ryzen 5600 / RTX 3060 Ti / Windows 11
Lamberto Vitali
Posts: 83
Joined: Fri Feb 25, 2022 12:21 am

Re: Run two tasks per GPU?

Post by Lamberto Vitali »

I thought of that, but Boinc tends to take more than its fair share. There is no way to throttle Boinc's GPU usage that I know of apart from by temperature.

Ignore this part, see next post:

Code: Select all

It may be a PCI-E lane limitation, not CPU, as the slowed cards are sharing a PCI-E v2 x1 lane.  I'm going to try them on a much faster computer and see if the same happens.  If it does, I'll buy these:

To split (without sharing a lane) an x4 into 2 of x2:
[url]https://www.ebay.co.uk/itm/334454067999[/url]

To split (without sharing a lane) an x16 into 4 of x4:
[url]https://www.ebay.co.uk/itm/393646739141[/url]

To convert the resulting M2 slots back to PCI-E for risers:
[url]https://www.ebay.co.uk/itm/325133059753[/url]

(These are the risers I already use, which share 4 cards from 1 lane):
[url]https://www.ebay.co.uk/itm/175305130369[/url]

If it is a lane limitation, it appears you can get 6Tflops worth (going by the specs of the actual card on techpowerup) through 1 v2 lane before it throttles, as a single 280X (4Tflops) is ok on 1 lane, but two at once they both go at 75%, so 1.5 cards worth.

Unfortunately risers don't do v3.  I need them as you just can't fit 10 cards in one motherboard.  And since GPUs are so much better at computing than CPUs, I want to just buy those and not spend the money on more computers.
Last edited by Lamberto Vitali on Wed Jun 08, 2022 2:42 pm, edited 4 times in total.
Lamberto Vitali
Posts: 83
Joined: Fri Feb 25, 2022 12:21 am

Re: Run two tasks per GPU?

Post by Lamberto Vitali »

Update - I was right in the first place. Two 280X cards CAN run shared on a single PCI-E v2 x1 lane, both going at full speed on Folding. The other CPU must have been too slow, which is odd as the usage was a fraction of a core to help each card. Maybe the delay for each bit of the CPU part of the work was enough to slow the cards down. So if you want to run many cards, just use the bitcoin quad risers, but make sure you have a decent CPU.

This CPU is too slow: N3700, scores 1169 on cpubenchmark.net
This CPU is fast enough: i5-8600K, scores 10321 on cpubenchmark.net
Lamberto Vitali
Posts: 83
Joined: Fri Feb 25, 2022 12:21 am

Re: Run two tasks per GPU?

Post by Lamberto Vitali »

Now it's not so slow anymore. Depends which cards are plugged in. One of them is being flaky and even dozes off completely. Maybe it was hogging the PCI-E lane trying to work? Or the Folding app was repeatedly sending it data to retry?
MeeLee
Posts: 1375
Joined: Tue Feb 19, 2019 10:16 pm

Re: Run two tasks per GPU?

Post by MeeLee »

Lamberto Vitali wrote: Wed Jun 08, 2022 2:39 pm Update - I was right in the first place. Two 280X cards CAN run shared on a single PCI-E v2 x1 lane, both going at full speed on Folding. The other CPU must have been too slow, which is odd as the usage was a fraction of a core to help each card. Maybe the delay for each bit of the CPU part of the work was enough to slow the cards down. So if you want to run many cards, just use the bitcoin quad risers, but make sure you have a decent CPU.

This CPU is too slow: N3700, scores 1169 on cpubenchmark.net
This CPU is fast enough: i5-8600K, scores 10321 on cpubenchmark.net
Your N2700 is fast enough to run 3 to 4x RTX 2060 gpus, so long you'll grant it at least a PCIE 3.0 x4 slot per gpu.
What's not fast enough is your PCIE connector.
An X1 slot is bottlenecking even a budget card like a GTX 1050 these days.
Lamberto Vitali
Posts: 83
Joined: Fri Feb 25, 2022 12:21 am

Re: Run two tasks per GPU?

Post by Lamberto Vitali »

The N3700 doesn't have that many lanes. It only has 4, I believe one is for SATA, one for network, 1 for USB, something like that. Only 1 is present at the only PCI Express slot.

You don't need anything like v3 x4. I've got two 280X cards running maxed out on folding right now, sharing a v2 x1 slot on an i5 machine. A GTX1050 is half the speed of a 280X.

High bandwidth is for games, not computation.
MeeLee
Posts: 1375
Joined: Tue Feb 19, 2019 10:16 pm

Re: Run two tasks per GPU?

Post by MeeLee »

Lamberto Vitali wrote: Wed Jun 08, 2022 8:57 pm The N3700 doesn't have that many lanes. It only has 4, I believe one is for SATA, one for network, 1 for USB, something like that. Only 1 is present at the only PCI Express slot.

You don't need anything like v3 x4. I've got two 280X cards running maxed out on folding right now, sharing a v2 x1 slot on an i5 machine. A GTX1050 is half the speed of a 280X.

High bandwidth is for games, not computation.
I think you should read up on the rest of the forum, before you make this statement.
PCIE x1 slots are for slow cards only. The 280x is several generations old, and according to UserBenchmark only 23% faster than a 1050, pretty much in the same league. That means a single one of them should already be limited in performance on an x1 slot (you won't know, unless you try an x4 slot).
Two of them on an x1 slot might work, but both GPUs will be severely PCIE bandwidth throttled.
You'll basically run those 2 GPUs at ~50-75% more power than a single one on an x4 slot, while gaining the same to 20% higher performance at best.
Not very efficient if you ask me.


I do apologize on the CPU, my CPU was a Pentium Gold G5400, which has 16 lanes, but otherwise is the same 2 core 4 thread CPU, with similar frequencies.
4PCIE lanes for sure is a limitation.
Sometimes Intel CPUs placed on the right motherboard, allow you to disable the IGP, and redirect the PCIE lanes to an x16 slot.
But this is not one of those apparently.
The CPU might be too slow to actually saturate the x4 slot on a high end GPU; so if you can get a PCIE x2 connection,
Lamberto Vitali
Posts: 83
Joined: Fri Feb 25, 2022 12:21 am

Re: Run two tasks per GPU?

Post by Lamberto Vitali »

It's over double the speed of the 1050:

1.862 TFlops: https://www.techpowerup.com/gpu-specs/g ... 1050.c2875
4.096 TFlops: https://www.techpowerup.com/gpu-specs/r ... 280x.c2398

I buy old stuff as it's a lot cheaper. Way more computations per dollar. The price of new cards is beyond a joke.

I know if it's running flat out and it is. And I can run three of them flat out on 1 slot too. Same speed for three as one, so I guess I must be right. No idea where you're getting your figures from, but you're simply wrong. If you were right then the three cards I have sharing a slot would not be getting three times as much work done as the one on its own slot. (These are all on an i5 8600K which easily manages)

The IGP does disable itself when a GPU is present. But with only 4 lanes I guess they're needed for SATA, network, USB. I think I read somewhere the MB only has 1 lane physically wired up to the socket anyway.
Post Reply