some sticky info for gpu

[James Heinrich]

I can't comment on how efficiently the F@H client uses GPU resources, but it runs very unobtrusively, I really can't tell if it's running on not based on system responsiveness (whether on desktop, watching a video or gaming it stays very well out of the way). Also it uses a trivial amount of CPU time (~2%).

Now I certainly understand mfaktc's use of CPU time, and I certainly don't mind giving it a core to play with, but it would be wonderful if it could emulate F@H's "invisibility" while running.

[mdettweiler]

Originally Posted by James Heinrich

I can't comment on how efficiently the F@H client uses GPU resources, but it runs very unobtrusively, I really can't tell if it's running on not based on system responsiveness (whether on desktop, watching a video or gaming it stays very well out of the way). Also it uses a trivial amount of CPU time (~2%).

Now I certainly understand mfaktc's use of CPU time, and I certainly don't mind giving it a core to play with, but it would be wonderful if it could emulate F@H's "invisibility" while running.

I know that BOINC is, by default, configured to pause crunching on GPUs whenever the computer is in use; perhaps Folding@Home does something similar.

[axn]

What happens if there is an onboard graphics card and it is used to drive the display -- will that be sluggish also (while the computation is happening on the discrete CUDA card)?

[mdettweiler]

Originally Posted by axn

What happens if there is an onboard graphics card and it is used to drive the display -- will that be sluggish also (while the computation is happening on the discrete CUDA card)?

No; the GPU and integrated graphics are entirely separate, so crunching on a GPU with a monitor on the integrated graphics should leave the comptuer fully usable.

Note that some BIOSes will autodetect which graphics system the monitor is on and automatically disable whichever is not in use. If that is the case, there will probably be a BIOS setting by which you can force it to keep both active.

[James Heinrich]

Originally Posted by mdettweiler

I know that BOINC is, by default, configured to pause crunching on GPUs whenever the computer is in use; perhaps Folding@Home does something similar.

It does not, at least as far as I can tell. It continues to use "idle" GPU time (although, as mentioned, I'm sure the priority scheduling for GPU is less nice than CPU scheduling). For example, while running mfaktc and F@H, the latter runs about 8x slower than running solo, but does continue to run.

[jasonp]

There is also Nvidia's driver to consider; CUDA programs can be set up to cue a huge amount of work to the card, but the driver will feed the card in little pieces and stop once in a while to let the card do graphics. If those little pieces each finish very quickly then the system should remain responsive, but CPU usage will go up as the driver needs CPU resources to keep the card fed.

Of course, if the pieces finish quickly on a GPU then odds are they could have finished quickly on the CPU as well.

[xilman]

Originally Posted by jasonp

Of course, if the pieces finish quickly on a GPU then odds are they could have finished quickly on the CPU as well.

Quite likely true but very likely irrelevant.

Assume that a piece takes 1ms on a GPU and 20ms on a CPU. Assume, further, that you want to run a million pieces. Where would you rather those pieces ran?

Well, that's how I design my GPU software, anyway. I'm at least as interested in latency as I am in throughput. Perhaps I'm strange that way.

Paul

[TheJudger]

Originally Posted by Uncwilly

Maybe we can use this thread to do the planning and organizing for the stickies.

Luigi (ET_) or Oliver (TheJudger) should be made a mod of this sub-forum.

I'm willing to help, too!
If the admins decide to do so: don't give me more rights than needed, please. If possible just for the related sticky thread.
Sometimes it is much better just to be a normal user.

Oliver