F@H - Multi-GPU Client/Core

[KE1HA]

Folding @ Home News

10/27/2006 Multi-gpu support coming soon
We are testing a new client/core (core v0.09) which allows one to use multiple GPUs in one box. We hope to release this next week.
.

Anbody know if this will still only take one Core?
.

 

[SiriusB]

Well Stanford said reducing the polling would be a simple matter. So I reckon the can get a pair of GPUs running on one core.

SiriusB

 

[FreeStream]

Well Stanford said reducing the polling would be a simple matter. So I reckon the can get a pair of GPUs running on one core.

SiriusB

Hell yea, more CPU time = win!

 

[Stan_Lite]

Well Stanford said reducing the polling would be a simple matter. So I reckon the can get a pair of GPUs running on one core.

SiriusB

Tried this last night.

Got the new core and started up a second GPU client in crossfire. It runs both clients - but very poorly. One was taking 13.5 minutes per frame and the other was taking 22.5 minutes. Even running no CPU clients made very little difference.

I had a read over on the Stanford forums and somebody there had tried it with 2 separate cards not in crossfire and it worked fine.

If you want to run 2 GPU clients in one rig, it would appear you'll have to do it separately - looks like crossfire is not supported

Stan

Edit: It looks like the polling issue still hasn't been addressed (according to the Stanford forums).

 

[rich99million]

In terms of the bandwidth requirements of the client this has been posted on the forums:

Okay, I finally got around to doing this. Intel 975X mainboard, E6400 processor

Project 2730 with board@3D clocks (650c/775m X1900XTX 512MB), minutes per completion frame:

x16: 6:02
x8: 6:08
x4: 6:51

The x4 slot on this board is via southbridge, but I don't have another board with an x4 via the northbridge I can test. Looks like minor differences in x8 vs x16. This is kinda expected since current download/upload rates to GPU is ~1GB/s, so we really aren't stressing the x8 bus which is theoritecal ~2GB/s, vs ~4GB/s for x16. x4 is in theory ~1GB/s, so we are stressing it, plus the added latency of the southbridge, and everthing else in the system.

It would be interesting to see others' measure on PCIe speed. If someone has on of the 1950 AGPs, it would be interesting to see the speed comparison against the x4 above. Sadly, we can't test with the same CPU to take that out of the equation, but it should give us ballpark.

from here

hopefully this should mean that an 8X AGP card will not suffer from bottlenecks due to the slower bus - hurrah!

 

[KE1HA]

I think I'm going to stick with one GPU per box for now, until they get this sorted out, unless they up the points on GPU WU's
.

 

[2bullish]

Even if you have 2 GPU's running in one box they will both work on only one WU, the gpu client does not support crossfire or multiple gpu's