85% GPU Usage

[kencode]

Code: Select all

Thanks for the info....now I see all 9 cards are throttled down. PPD has gone from 200k to 140k.

Maybe the moderator can jump in here.

KRC

[bruce]

Off topic:

I have similar results to the above poster. I run 3 GPU's in one system. 2 gtx 480, and a single gtx560. . .

I don't know how you are managing to fold with those cards overclocked? . . .

Off topic discussion moved here.

Let's focus on the WUs and the way they load the GPUs, not how to build a system. That discussion belongs in the hardware forum.

[{RaW}Eagle1]

Not a setting problem. It's a larger work unit, and the GPU processors are not the limiting factor for these WUs. Feeding the data fast enough to keep the GPU busy seems to be the limiting factor with these WUs.

I responded to a post Bruce put in on this issue earlier with regard to what one might expect to see when changing the shader clock speed up from where it was originally (with the cards running at 86% load). Bruce suggested that if you were to overclock the card you would expect to see the load on the card drop as the GPU would spend more time essentilly waiting for work. I have tested this and it is not the case:

Original setup
GTX480 SLI @ 650|1300|1900
Load at 86%, 14000 PPD per card (approx)

Test setup
GTX480 SLI @ 810|1620|1900
Load remained at 86%, 17500 PPD per card (approx)

The strange thing for me is that there is no drop in the load on the cards at their higher clock setting. Their productivity goes up proportional to the increase in clock speed, yet the utilisation of the GPU remains constant. . . This was the evidence that led me to theorise that it was not PCIe saturation / bandwidth that was the problem as there is clearly the ability for the GPU to be fed with more data, but more likely a latency issue.

How crucial is the clock speed of other system components to these larger work units? Could it be graphics / system memory speed that limits output? Is everyone experiencing the same 86% load or does the load percentage vary from system to system / card to card?

For the record my GRAM is as follows: GDDR5 @ 1900 MHz (Width: 384 bits).

Thanks for continued support.

Eagle

[Jesse_V]

How crucial is the clock speed of other system components to these larger work units? Could it be graphics / system memory speed that limits output?

I've heard that memory clocks have very little impact. But maybe it's PCIe bandwidth. Check the first page of this thread.

[bruce]

It's certainly a valid question ... and easy to test. Change the memory clock and see what happens. With small proteins, it never made enough of a difference for anybody to care but we're not dealing with small proteins any more.

The other think that's worth testing is how important is PCIe speed. On most systems, the PCI can be reduced from 16x to 8x fairly easily.

One thing at a time, though.