Understanding and interpreting GPU utilisation

[akarypid]

Hi all,

Maybe someone more knowledgeable than myself can explain or point towards resources on how GPU utilization is reported? When I see a video of someone showing a benchmark, what does that number really mean?

For example, assume I see this readout:

FPS: 60
GPU%: 80
CPU%: 50

... and then later:

FPS: 60
GPU%: 100
CPU%: 50

I'd assume that the first area is easier to render than the second (maybe lighting, shadows, whatever...) and therefore the card is "underutilized".

But what does 80% mean?

Trivially one might think that if the compute capacity of the card is 10TFLOP per second and the label displays the utilization for the last 1 second (i.e. values are updated 1/sec), then one could say that the card executed 8TFLOP in that second?

But then what if the card has lowered its frequency to save power, or because the user has simply down-clocked it, or if it's simply thermal-throttling? Is that taken into account?

For example, in CPUs this happens (see this link from Microsoft technet) and if a CPU has a single core with max frequency 2GHz but is running at 1GHz and is fully busy at the time, it will register 100% utilization even though it's really 50% (clock is at half speed).

 

[LegendaryDread]

Trivially one might think that if the compute capacity of the card is 10TFLOP per second and the label displays the utilization for the last 1 second (i.e. values are updated 1/sec), then one could say that the card executed 8TFLOP in that second?

You're correct - and it's not measured any differently from CPU usage. However, there's more to factor in to diagnose 3D performance, such as VRAM usage, or memory bandwidth.

At a given load, usage directly correlates with performance.

As for downclocking, yes, that's taken into account by [I assume you're talking about] RTSS Overlay. By default, utilization/power/voltage/thermal limits cause downclocking, and as such the GPU becomes less powerful, meaning an increase in GPU usage.

 

[akarypid]

You're correct - and it's not measured any differently from CPU usage. However, there's more to factor in to diagnose 3D performance, such as VRAM usage, or memory bandwidth.

At a given load, usage directly correlates with performance.

As for downclocking, yes, that's taken into account by [I assume you're talking about] RTSS Overlay. By default, utilization/power/voltage/thermal limits cause downclocking, and as such the GPU becomes less powerful, meaning an increase in GPU usage.

Yes exactly, things like RTSS.

So, when comparing two cards with similar TFLOPs like the 390X and 480, if one makes sure they are never throttling (disable any power saving, make sure cooling is adequate, don't overclock, etc) then a less than 100% utilisation would mean that the game is CPU bound?

I mean, as long as the CPU can keep feeding the GPU for work, the FPS would go up until 100% GPU usage is hit, right?

EDIT: reason why I mention cards with similar TFLOPs is because I was thinking you can see the impact of architectural improvements and other design choices (ROPs/TMUs) as long as you make sure the cards are not throttling and CPU is not an issue. Any other factor to consider?

 

[eddyr]

I've tried to look this up before but never found a clear answer on what GPU % is and how it is measured. I suspect it is measuring how busy the hardware that is scheduling and dispatching work is reporting itself to be, rather than execution.