Core 17 eats 1 cpu core with Nvidia GPU?

[jeano]

Hi Guys

I have noticed that Core 17 uses 1 CPU core on my machine. As a result of this my CPU folding now takes twice as long (I have a dual core CPU) and my GPU is folding slower.

It looks like my CPU is a bottleneck on Core 17, this was not the case with Core 15.

Is this correct or am I the only one experiencing this? IIRC only AMD GPU's use 1 CPU core to fold or could this be due to the fact that core 17 is using opencl and not CUDA?

EDIT: The CPU is E6300 @2.8ghz Wolfdale (the 45nm and not the 65nm running @ 1.8ghz).

[Joe_H]

Yes, this is a change with Core 17 due to the way nVidia's drivers handle OpenCL commands compared to CUDA commands used by Core 15. WIth Core 17 persons are reporting that the AMD drivers use only a portion of a CPU core to process data and the OpenCL programming.

[7im]

Remove the "advanced" setting and you won't get core_17 work units.

[jeano]

Remove the "advanced" setting and you won't get core_17 work units.

That’s only a short term band aid. Your suggestion crossed my mind when I first noticed it.

@Joe_H
Is there room for big improvements with Nvidia opencl calls by the time Core 15 is stopped?

[GreyWhiskers]

I'd suggest that you take a close look at the PPD performance of both your GPU and your SMP folding slots. Then optimize the resources assigned to the various slots to maximize the total return. While Core 15 may allow the CPU folding more resources than Core 17, your overall points production is probably going to be less.

In my case, even on the Core i7 2600K desktop I have, the GPU slots get so many more ppd than the SMP slots that the sacrifice of CPU cores/threads to favor the GPU folding is worth it (IMHO). If overall science production, as measured by the aggregate ppd of the various work units you complete, is the thing you are trying to optimize, take the hit on the SMP slots - drop back to a single core Uniprocessor - let Core 17 have all the resources it needs.

Of course, the relative points production is something that isn't static. Things could change tomorrow.

[PantherX]

The feature of Nvidia GPUs using 1 CPU per GPU is dependent on the driver. There are two methods used for OpenCL, one uses very little CPU (AMD Drivers use this method) while the other uses 1 CPU (Nvidia is using this method). For a technical answer, please read this post (viewtopic.php?p=242466#p242466).

[k1wi]

The other thing to note is even if you allocate one less smp slot to your CPU, if your CPU does hyperthreading, the performance drop should be relatively insignificant...

[Joe_H]

@Joe_H
Is there room for big improvements with Nvidia opencl calls by the time Core 15 is stopped?

PantherX has provided some information on the differences between the core utilization between nVidia and AMD. The most I can add is that changes in the driver are ultimately up to nVidia. They may do so, but at the moment are pushing CUDA for GPU computational use. That gives them a proprietary edge as applications written for CUDA will not run on their competition's cards. PG attempts to stay vendor neutral, and the group going to using OpenCL simplifies software development as the same code base can be used for both.

[folding_hoomer]

I'd suggest that you take a close look at the PPD performance of both your GPU and your SMP folding slots. Then optimize the resources assigned to the various slots to maximize the total return. While Core 15 may allow the CPU folding more resources than Core 17, your overall points production is probably going to be less.

In my case, even on the Core i7 2600K desktop I have, the GPU slots get so many more ppd than the SMP slots that the sacrifice of CPU cores/threads to favor the GPU folding is worth it (IMHO). If overall science production, as measured by the aggregate ppd of the various work units you complete, is the thing you are trying to optimize, take the hit on the SMP slots - drop back to a single core Uniprocessor - let Core 17 have all the resources it needs.

Of course, the relative points production is something that isn't static. Things could change tomorrow.

Sorry, but i think this is not a rule in general - it belongs to the (folding-)power of your GPU.
As far as i know:

If you use an older AMD-GPU like 58xx, 6870 and 69xx your GPU might profit by using Core17 and appropriate WU´s (5870 tested by myself) - if your GPu has less power use the older Core16 to maximize your points.
But - most "new" AMD CGN-GPU´s definitely profit while folding under Core17.

The same belongs to nVidia:
If you use GTX 460 (470, 480, 560 and up, 650 and up) you will profit by using Core17, if your GPU has less power use Core15.

But - there might be some WU´s using Core17 where you need a faster GPU than a GTX460 to get more points . . .

So - it´s your decision what to do, but - choose wisely

[bruce]

A video driver which uses 1-cpu gives somewhat better video performance than one which uses very little cpu. When people measure a GPUs game performance, the computer almost always has idle CPUs. It's a question for [NVidia/AMD/Intel/etc.] to decide which sells more GPUs: Video games or analysis by GPUs -- as well as a question of whether there is a driver based compromise that can be called "optimized" by everyone.

Contact your GPU manufacturer.

[GreyWhiskers]

That's why I said check. Each of our systems is different. Be objective in what your goal is, then set up your system to give it to you. Go back and re-check peridocally if the distribution of projects changes.

You may get better ppd with your SMP and core 15 setup, or you may get better aggregate performance from the GPU core 17 with less dedicated to SMP. Your setup differs from any of mine. There is no a priori universal answer.

YMMV

[bruce]

... and if you get a mixture of assignments for more than one FahCore, you may need to optimize for different combinations of assignments.

Stanford's benchmarking process assumes only one WU is being processed. The choices made by adjusting the slider was an attempt to avoid some of the worst choices, but it's terribly complex if you consider all possible combinations on all possible systems so software can't optimize YOUR system, it can only make automatic choices that avoid really some bad choices. You can possibly do somewhat better yourself.

[Jesse_V]

The goal of the "Full" setting on the slider is to get 90% of the performance in 90% of the machines. Obviously FahCore 17 changes things a bit, hence https://fah-web.stanford.edu/projects/F ... icket/1013

[TonyStewart14]

I had this same experience. I tried setting client-type to advanced for my GT 620 GPU on v 7.2.9 with a Pentium D processor. The PPD went from ~5.5K to ~2K, although there seems to be a small QRB with core 17. Interestingly, the CPU is the bottleneck as it can't use more than the equivalent of one core (50%) on the dual core processor, although from Task Manager it appears that both cores are working mostly the same amount and I am folding on GPU only.

[n_w95482]

I had this same experience. I tried setting client-type to advanced for my GT 620 GPU on v 7.2.9 with a Pentium D processor. The PPD went from ~5.5K to ~2K, although there seems to be a small QRB with core 17. Interestingly, the CPU is the bottleneck as it can't use more than the equivalent of one core (50%) on the dual core processor, although from Task Manager it appears that both cores are working mostly the same amount and I am folding on GPU only.

I doubt a Pentium D is bottlenecking a GT 620, it's just simply pegging out a thread (yay Nvidia OpenCL!). There's definitely a point of diminshing returns with running core 17 vs 15. In my testing, an overclocked GTX 460 1 GB comes close to breaking even, but not quite. I have a GTX 570 running in my dual-core Opteron machine (Athlon 64-era) and it runs at full load with a bit more PPD on core 17 than 15. I've stopped CPU folding on it due to core 17 keeping half of the CPU busy.

I'd definitely recommend removing the client-type option. Not only will you get more PPD as long as core 15 WUs are available, it'll let your CPU idle which should cut back quite a bit on power and heat.