Actually, it doesnt completely load the GPU with 3 instances.
But 94~97% is enough for me.
And it's lag free!

[QUOTE=Karl M Johnson;254113]Actually, it doesnt completely load the GPU with 3 instances.
But 94~97% is enough for me.
And it's lag free![/QUOTE]
I was actually thinking that it might be worth asking your cpu to do a bit more work in order to have less candidates to pass on to the GPU. Especially if your GPU is not saturated I suspect that sieveprimes will be at minimum which is not very optimal unless there is free time on the GPU.

What programs are people using to see GPU load?

I can't see to find anything in the nvidia app installed to my PC in the driver bundle.

-- Craig

The CPU is Q6600@3.0 Ghz.

GPU-Z, RivaTuner and it's clones MSI Afterburner, EVGA Precision.

or if you have a gigabyte card, GIGABYTE OC GURU ( under windows only it seems)

[QUOTE=Karl M Johnson;254159]The CPU is Q6600@3.0 Ghz.

GPU-Z, RivaTuner and it's clones MSI Afterburner, EVGA Precision.[/QUOTE]

Cool. GPU-Z wins for me. Rivatuner was what I was using - but wasn't working for me (it runs - but I can't find any usage stats, I'm sure it was there previously).

-- Craig

On Linux it can be done without 3rd party utilities:
[CODE]nvidia-smi -a[/CODE]

Oliver

[QUOTE=TheJudger;254172]On Linux it can be done without 3rd party utilities:
[CODE]nvidia-smi -a[/CODE]Oliver[/QUOTE]Thanks. I never knew that and have just learned something valuable!

Paul

Hi Paul,

you can use it to enable or disable ECC on current Tesla or Quadro, too.
It reports ECC errors aswell. :smile:
AFAIK it will report free/used memory in feature versions, there was something on the Nvidia CUDA forum.

What is doesn't do is temperature readings on Tesla M20x0 (while it works on C20x0)...

Oliver

Win7 Gadgets
 

[QUOTE=nucleon;254156]What programs are people using to see GPU load?
[/QUOTE]

There are some nice gadgets for Win 7 out there: see [URL]http://blog.orbmu2k.de/[/URL]
I use the GPU Observer Gadget, see attached (screen includes CPU gadget).
It also supports Fahrenheit.

The page is in German though.

Polling vs. Event-based wait-for-GPU
 

I have this laptop with a Core i7 M620 (~3GHz), and a NV QuadroFX 880M (core 550MHz/mem 790/ shader 1210), running Win7/64

mfactc 0.13 64-bit on that HW does TF of a M100xxxxxx 68 69 about 3 times the speed of prime95, a clear advantage.

However, the CPU seems to be so much faster than the GPU, that even with SievePrimes=100000, the GPU stays at 99% load (just one mfactc instance).

I then used Throttlestop to reduce the CPU-clock-speed. Even at 1.2 GHz, the GPU was @ 99%, with only occasional drops to 94%. The FC throughput just dropped from 12.62M/s to 12.49M/s (~1% slower when CPU is 60% slower).

With the CPU running at 3GHz this would mean that about 1.8 GHz are wasted on the busy-polling while it could drive a prime95 thread at about half-speed.

Is it really that uncommon to have a decent CPU with a rather weak GPU?

I just hate wasting cycles and therefore wanted to ask for raising the priority of an event-based implementation of the GPU synchronisation.