SMP plus GPU?

[yahavbr]

Hi, this might sound silly for some so sorry in advance!

I'm successfully running SMP client on my Windows 7 with I5 quad core. It's running fine, taking 100% CPU as expected.

I also have graphic card in the machine with half giga memory of its own.... I wonder if I can run GPU client in addition to the SMP client?
Won't it make CPU over the limit or something?

Thanks in advance!

[PantherX]

It depends on the type of GPU which you haven't stated:
ATI GPU -> You might need -smp 3 plus environment variables
Non-Fermi GPU -> Not an issue with GPU2 or GPU3 BETA
Fermi GPU -> Might be an issue with the latest Projects as they need more CPU Cycles so -smp 3

[Racer43]

In the past, I ran an ATI 5670 with environmental variables working the card between 97% and 99% usage and SMP on my system listed below in the past. The ATI card uses between 2% (p573x) and 25% (p574x) of a single cpu. With the smallest SMP WU (421 point I believe) and 75% average of total cpu usage, I'm at 120 ppd on SMP and in severe danger of not meeting the final deadline (especially when the machine is used for HTPC); at 97% average cpu usage, I'm at 270 ppd and can make the preferred deadline. However, with SMP at 99% cpu usage, I'm at 320 ppd and well within the preferred deadline. With the machine you have, one could make the preferred deadline for SMP with ease and still run GPU with ATI or nVidia cards.

But a word of caution. I no longer run SMP with GPU on my particular machine due to how long it takes to return the SMP WU; the next step in that specific WU run cannot start until my results are returned, so pointwise I may be ahead, but sciencewise I am slowing things down. I run GPU with uni; it frustrates me at times to see unused cpu cycles, but in my mind, I am contributing to science as fast as I can (squirrels can only run so far, so fast).

[Zagen30]

I run GPU with uni; it frustrates me at times to see unused cpu cycles, but in my mind, I am contributing to science as fast as I can (squirrels can only run so far, so fast).

Are you running multiple copies of the uni client? If you are, then you shouldn't have any unused CPU cycles, unless you're deliberately keeping one of your cores free for the 5670.

[Racer43]

If one gets two b4 at the same time, then it slows my GPU down by almost 200% (my GPU is my points workhorse). Since I cannot control what is assigned, I just run one uni.

[bruce]

If one gets two b4 at the same time, then it slows my GPU down by almost 200% (my GPU is my points workhorse). Since I cannot control what is assigned, I just run one uni.

Even if you set the environment variables, an ATI GPU is going to take enough processing to disrupt the SMP client.

Check the idle CPU time with one uniprocessor client plus one GPU client. If it's small, then that is the best configuration for your hardware. If it's in the 40% range, then you can consider two uniprocessor clients plus the GPU. In either case, follow the recommended settings for priority: Uniprocessor or SMP set to lowest possible. GPU client set slightly higher.

[uncle fuzzy]

You still haven't mentioned which GPU you have. If it's nVidia, I say go for it. If it's ATi, I say try it.

Worst case would be an ATi that wanted to play core-hog. That might lead you to running -smp 3 and adjusting the EVs to maximize the output of both clients.

Many of the slower duals will no longer make deadlines with the current SMP WUs, even without a GPU client taking core time away from it. I have not heard reports of any quads having this problem. Using the PPD=amount of science done rule-of-thumb, the GPUs do more than the SMP. Most current GPUs (NV) will get 1 1/2 to 2x the points/science of the SMP. Even if you have to back down to -smp 3, the total machine will still pump out more than double the work.

I have no ATi, but the gain will be much lower.

I have 5 Intel quads with 2 or more NV GPUs in each. The only one that can't run the SMP on all cores is the one with 3x GTX460s.