Understanding GPU load numbers - what does it actually measure?

junky

Member
Hi all!

Long time user/dreamer

In the wake of DX12 and DX12 benchmarks and all the discussions and in-depth info around them, I started to question my understanding of what do the GPU measurements actually say:
1. GPU Core Load [%]
2. GPU D3D Usage [%]
3. GPU Bus Load [%]

Previously I thought that if the core load is 99-100% then it probably means that the GPU activity is almost at full capacity, but now I'm not sure what it means. If before, with DX11, not all the GPU capabilities/potential were utilized, then how come you got 100% load?
If AMD GPUs weren't fully utilized (when supposed too, without CPU limitations) and now they are thanks to DX12, better programming/multi-threading/ACEs/more, then what is the difference?

Could someone help me with that?

Thanks
 
"GPU D3D Usage" is the usage of the Direct3D subsystem only, so it doesn't cover usage of GPGPU computing, or other subsystems.
"GPU Bus Load" is the usage of PCIe interface only.
As for the GPU usage itself, GPU vendors don't say how exactly is this determined. I suppose it's similar to CPUs - the CPU can be loaded at 100% (probably all instruction queues full), but there still can be several resources in the CPU idle.
 
Martin said:
"GPU D3D Usage" is the usage of the Direct3D subsystem only, so it doesn't cover usage of GPGPU computing, or other subsystems.
"GPU Bus Load" is the usage of PCIe interface only.
As for the GPU usage itself, GPU vendors don't say how exactly is this determined. I suppose it's similar to CPUs - the CPU can be loaded at 100% (probably all instruction queues full), but there still can be several resources in the CPU idle.

Thanks Martin. So it is all based on some not fully clear sensors? So, in fact, when I see 100% GPU usage it doesn't mean there isn't place for higher GPU output?

Also, what do you mean by Direct3D subsystem? the rasterization engine without the compute engines?
 
I'm sorry, but I don't know more details what exactly (on chip level) those usages reflect. You might try to ask NVIDIA for a more precise explanation.
The D3D Usage is based on certain internal D3D statistics that are also not well described.
 
Martin said:
I'm sorry, but I don't know more details what exactly (on chip level) those usages reflect. You might try to ask NVIDIA for a more precise explanation.
The D3D Usage is based on certain internal D3D statistics that are also not well described.

:\
Can you please refer me to the right NV contact? I've tried many times to contact many companies with such questions and barely got replied

Anyway, many thanks!
 
I'll try to ask.
What exact questions do you have regarding the GPU Core Load ?
 
Martin said:
I'll try to ask.
What exact questions do you have regarding the GPU Core Load ?

I don't have the knowledge of what's going on inside, but in my simple words - I want to know what this load includes and what it doesn't. Otherwise, what do we really learn from it? Specifically, I wanted to use this tool to know when I'm GPU limited and if/which ACEs equivalent were in use

Many thanks!
 
Back
Top