What 'GPU Usage' really means
17 Feb 2014 22:11:16 UTC
Topic 197386
(moderation:
The reason why we run two or more wu's together is that we want a high gpu usage. For my main system a setting of 3 wu's per gpu works best, gives ~96% for the AMD-card.
But when I switch over to Milkyway, I hear the fans spinning up, see the power meter climbing, but the headroom for the gpu usage is very small, it increases from 96% to 99%. This alone cannot explain the higher power consumption and the fans spinning up.
So can anybody explain, what gpu usage really means, which parts of the gpu are included (shaders, textures etc) and why the temps can go up 15 degrees by nearly the same usage?
Same thing happens when I switch my nvidia cards to gpugrid.
Cheers, Alex
What 'GPU Usage' really means
)
There's a major difference between E@H and MW tasks. MW tasks utilize double-precision and these tend to load the GPU execution units much higher. I had the same experience when I switched to MW and when running those units, my GPUs are ~10 C hotter than running 2 E@H tasks.
To answer your questions about "GPU load" - it's a good question, but I believe nobody can give you a precise answer except a few AMD engineers. This value you see (depending on the type of tool used to read it) is based on some internal calculations either by the GPU drivers, or internal GPU registers. But the exact meaning of this values is not disclosed to public.
I believe it's similar to "CPU load" - you can have the CPU 'loaded' at 100% and still a lot of CPU units are not utilized to the max. Or you can run a special-crafted code which will load your CPU close to its power/thermal limits.
Most probably the 'load' determines just if/how busy the CPU/GPU queue is, but not all execution (and other) units.
-----
I agree: if the GPU is
)
I agree: if the GPU is reported to be utilized 100% this means it's not free for any other tasks. But how efficiently the currently running program can utilize the execution ressources is a totally different question.
MrS
Scanning for our furry friends since Jan 2002
I was contemplating on
)
I was contemplating on getting a GPU with better "double precision" performance. This is the only thread mentioning anything about it, so E@H doesn't even use "double precision" in the cuda apps? Is this correct?
Saw on this older post by
)
Saw on this older post by Bikeman, that 100% GPU usage in this app means actually 75%.
http://einsteinathome.org/node/195578&nowrap=true#109696
If this is still correct for all the updates too, that would explain a difference to the MW tasks.
I think this is from a era
)
I think this is from a era where the BRP CUDA app was still pretty new and from the beginning not all of the computations were done on the GPU but quite a lot of them were run on the CPU. I think this might have been the first release where all of the computations were done on the GPU with the CPU only doing the setup of instructions to the GPU and telling it how to process the work, so 100% refers to all computations move to the GPU and 75% is probably the observed load on the test/development system's GPU.
RE: I was contemplating on
)
E@H is usgin pure single precision for the GPU apps. It's not that "they don't even use DP", but rather that "they can get by without it". Using DP is not better per se, these calculations require significantly more energy to perform and any hardware becomes slower in DP mode, no matter how good it is compared to other hardware.
The only 2 BOINC projects I know use DP are Milkyway and some sub-project of PrimeGrid. If you like Milkyway get a Tahiti based AMD, otherwise don't bother with the super-expensive nVidias with decent DP performance (Titan, Titan Black).
MrS
Scanning for our furry friends since Jan 2002
I've got a first cuda6 wu
)
I've got a first cuda6 wu @GPUGRID. Running one of the beta wu's there I see 78% gpu load, 300MB dedicated mem, 39MB dynamic mem, 22% mem controller load and up to 75% TDP power load.
Lower gpu-load but more power?
I posted this question @ TechPowerUp. If someone is interested to follow the answers here is the link
http://www.techpowerup.com/forums/threads/what-gpu-usage-means.198916/
RE: Lower gpu-load but more
)
Sure. See the last lines in Mumaks initial answer. GPU load is probably defined by "time being occupied with something", independent of how intense that workload actually is. It could be just slow memory operations, which will hardly need any ALUs at all, or it could be some densly packed number-crunching. Or anything in between.
MrS
Scanning for our furry friends since Jan 2002
It also needs to be said,
)
It also needs to be said, that reporting of GPU load can be quite different between GPU families. Some measure a certain level of GPU activity over a sampling period, others the level of SCLK activity, etc.
Though exact implementation is hidden in the internal GPU logic/SCU firmware...
-----