GPU slows down CPU folding?

[RMouse]

I have always folded with my computer using both the CPU and GPU. I thought that this was the way to generate the most output for the Pande Group. Folding this way on my Dell laptop allowed me to usually get between 200 and 300 points per day. Now however, by sheer coincidence, I started a folding session only using my CPU. I was shocked to see that the completion times on my WU dropped by half and that my points per day shot up to an unimaginable 1,000!!! I never once got such astonishing performance in almost 1 year of folding. WU's that used to take over 24 hours to finish were now being projected to finish around 12 hours.

Looking at the task manager I see now that one core gets 100% of my CPU power (the performance tab shows both cores nailed up at 100%). When I have the GPU folding, there are two cores, each getting 50% CPU power.

I assume that the GPU's performance is not visible anywhere on the task manager.

I'm left wondering why my folding performance is so poor with the GPU engaged.

I'm folding on a Dell Latitude E6400 with NVS 160M video card.

Any ideas??

[Jesse_V]

I assume that the GPU's performance is not visible anywhere on the task manager.

It's not. You have to use third-party programs like GPU-Z to get GPU performance info.

What GPU?

[bruce]

When you fold with a GPU, a certain amount of CPU processing is required to send and receive work to/from the GPU. The amount of CPU time is regulated entirely by the way Nvidia/AMD writes their drivers. For a long time, the ATI/AMD drivers required 100% of a CPU core and the NVidia drivers required almost zero. Thus any CPU with an NVidia GPU would be able to allocate all of the CPUs to SMP while the same CPU with an ATI/AMD GPU would be able to allocate only (N-1) of the CPU cores to SMP.

Add to that, the fact that SMP typically runs only as fast as the SLOWEST CPU thread and your AMD GPU will cut your SMP production in half, as you've observed.

The GPU's performance is not shown unless you download one of the GPU monitoring utility, but the CPU support processing is shown, being allocated to a program called FahCore_1x (where typically x=1,5,6). The CPU processing allocated to SMP processing is shown as FahCore_7x or _Ax.

The Latitude E6400 has a Core 2 Duo CPU, which means it can run SMP with two cores or it can devote one core to a uniprocessor task and the other to supporting the GPU. The NVS 160M GPU has only 8 CUDA cores which means it will produce less science than can be produced by allocating the second CPU to actually folding rather than supporting the GPU. I would immediately shut down GPU folding entirely on that machine.

[RMouse]

Thanks very much for the advice. I wish I noticed this earlier but oh well. From now on, both cores go to the CPU.

HUGE thanks!