smp vs gpu2 in terms of scientific benefit

[eliot1785]

Is the GPU2 considered to be much more valuable than the SMP client in terms of scientific benefit? I'm wondering this because it seems like the GPU2 offers vastly more PPD than the SMP client, at least if you compare top-of-the-line GPU's with top-of-the-line CPU's.

The GPU clients are benched around 1500 PPD and the SMP clients are benched around 1760 PPD, but the results in the wild seem to be much higher for GPU than the benchmark (generally above 4000PPD if you are using at least the 8800 series, from this chart: viewtopic.php?f=43&t=3193), perhaps because the GPUs being used are faster than the benchmark machine.

I'd like to know this for the sake of deciding between the SMP and GPU2 clients in the future, although I guess nothing beats a personal benchmark.

[eliot1785]

I just noticed the following so perhaps it partly answers my question for me:

"I also added the [Preliminary] to the thread title as this is still VERY early in the development process, and PPD IS going to change, probably many times."

[Leoslocks]

I'd like to know this for the sake of deciding between the SMP and GPU2 clients in the future, although I guess nothing beats a personal benchmark.

I thought everyone with a Dual or Quad was using both GPU2 and SMP clients? I hope to add a second GPU2 client to a dedicated Q9450 soon. FX Video has an 8800GTS for $109 including shipping.
[Edit]Oops, they are out of stock, Again.

[crodgers]

I currently run an SMP client on a Q6600 (Vista Home premium, 3GB). Can I also run a GPU client on the same machine (nVidia 8500GT)? Or do I have to choose between the two clients?

[ExtraFloppy]

I've chosen to run GPU2 clients only as the GPU PPD is so much greater I want to maximize the CPU's role (XP/E6600 cpu/8800gt x3), however the Vista CPU loading is much lower - you could probably get away with running 1 core of the cpu for several gpus and the other 3 cores for the SMP. I run two 8800gt's off of one of my CPU cores without any GPU PPD impact. There is a delta with the third card, but it overclocks better then the other two so I gave it a dedicated cpu core. I typicaly see 5300/5000/5000 ppd from the system.

[crodgers]

Wow! So, from what I've gathered, you need one CPU for every GPU you want to run. I am currently getting about 1700 ppd running SMP on all 4 CPUs. if I tell SMP to only use 3 CPUs, I can configure a GPU client to use the 4th CPU to assist the GPU? Interesting... I'll have to think about it... Do GPUs tend to heat up more than CPUs?

Also, where can I find tools to look at my past FAh logs and give me historic ppd by WU and average time per frame by WU? Or do people do that themselves? I am trying to get some baseline stats to compare to so I can see if/when i have a drop in performance on the box. i can also look at the stats to compare with and without GPU client...

[ExtraFloppy]

Not sure if you need the full use of a dedicated cpu core per GPU, but I bet it depends on the power of the gpu and cpu. With some of the GPU's out there delivering in excess of 7,000 ppd, getting an extra 500 ppd for the spliting a cpu core is rather moot. You would be better off getting the maximum capacity of the gpu. I'll try running the three 8800gt's off of a single cpu core tonight to see if there is a performance impact or not.

If you look here there are some gpu logs:

viewtopic.php?f=43&t=3193&start=510

For comparisons, that might be a little tricky, as CPU's and GPU's do not work on the same WU's. Each is assigned to the type of proccessor that best handles the problem at hand; cpu, smp, gpu, or ps3.

As far as heat. . .yes. LOTS of heat. The wu's are particuarly fond of shader overclocking, and with 112 cores chugging away constantly, even with manualy setting card fans to 85% core temps each maintain 75-80c. With high end cards featuring 240 proccessor cores, heat and power can be issues.

[rickduff]

If you are running XP, you pretty much need to dedicate a core to each GPU client you want to run (assuming you run more than one). If you are running Vista you can pretty much still use that core (Vista takes much less CPU time to keep the GPU fed and happy). That was pretty much the driving factor to using Vista on my folding machine. And as far as SMP vs GPU, you do both. My system for example has a 8800GTS and a Q9300 processor (both slightly overclocked) and with the good WU's (the ones with the high points for NV) and the new A2 core, I am seeing about 4800 ppd from the GPU and 5300 ppd with the SMP.

To get those kinds of numbers on the SMP I ended up running linux in two VMware instances. The Windows SMP client can't do anywhere close to those numbers. If VMware would let me use all 4 cores in one instance, I would switch to only using one instance.

Rick..

[eliot1785]

@rickduff, You have the a2 core on Vista? I assume for beta testing?

[Flathead74]

@rickduff, You have the a2 core on Vista? I assume for beta testing?

check here:

new A2 core released--on advanced methods

[eliot1785]

@Flathead74, the post you linked to mentions that it is for Linux and OSX. However, it looks like rickduff has the A2 core on a Vista machine, if I am reading his post correctly.

[rickduff]

Indirectly - I am running Linux under a couple VMs under Vista. Sorry bout the confusion. I did this so I could get maximum value out of the quad core. I tried both the Windows SMP client a 4 single clients and the best performance by far was with 2 VMs running the linux SMP client.

Rick..

[bollix47]

Indirectly - I am running Linux under a couple VMs under Vista. Sorry bout the confusion. I did this so I could get maximum value out of the quad core. I tried both the Windows SMP client a 4 single clients and the best performance by far was with 2 VMs running the linux SMP client.

Rick..

What distro and version of linux are you using?

[ExtraFloppy]

Results are in: After running two hours with 3x8800gt cards on a single core of a E6600 using XP (second core at 0% use) . . . .fahmon says. . . 15,310.18 ppd. No loss, XP's high cpu loading is not effecting the gpu.

[crodgers]

I tried both the Windows SMP client a 4 single clients and the best performance by far was with 2 VMs running the linux SMP client.

I have not used VMware yet, but I understand the concept - virtual machine = software running an environment that you load another OS into and use just like another PC - all running within your existing PC. Do you have links to some instructions that show how to set it up on an intel quad core running Vista - from scratch? What linux are you running in VM? Umbuntu? RH? SLES? Did you use a pre-configured VM package with VM player? Or build it yourself? If I can get more throughput from my existing box (without burning up the house), I would be game.

You mentioned that the SMP under Vista does not get nearly the amount of ppd as 2 X VM linux. If I switch to the 2 X VM linux setup,
* will the machine run hotter than SMP Vista?
* can I still try to run a GPU client - or will that just take away from the VM clients?
* if I want to try a GPU client, could it run in the VM as well?
* can I still run windows tasks periodically (rip a CD, check out forums like this, etc)?
* is it difficult to shut off the VM clients when the PC needs to be rebooted (Vista updates, hurricane, etc)?