How is GPU Usage/Load actually calculated/measured?

[Edgar R. Zakarian]

How do programs actually measure/calucate this? you know 0-100% usage. 
How does it know? like number of active transistors?

[porina]

It's pretty stupid actually. More a time load. This is especially important if you have a HT/SMT enabled CPU. Assuming workload is balanced, when you hit 50% you're effectively using all the "real" cores, and the region between 50-100% is the small extra benefit from HT/SMT.

 

Edit: doh! I misread the question as CPU not GPU. I'm guessing it is similar though.

[Hairless Monkey Boy]

As an educated guess: It seems like a simple ratio of active vs idle time over the sample period.

 

It has nothing to do with how many of the total transistors are being utilized. CPU's are composed of many sub-sections that each handle a particular type of calculation/workload. So measuring utilization in that way would not be useful.

[igormp]

There are MANY different ways to calculate load. You can see how many units are currently in use over the total number of units, you can see how much bandwidth is currently being used over the theoretical maximum, or use the number of gates, etc etc.

 

There are even some stupid ways that some people use , such as comparing the base clock vs the current clock speeds (such as MS with CPUs).

 

In the case of AMD on Linux with AMDGPU, here's how it's done:

Quote

The amdgpu driver provides a sysfs API for reading how busy the GPU is as a percentage. The file gpu_busy_percent is used for this. The SMU firmware computes a percentage of load based on the aggregate activity level in the IP cores.

Basically the firmware in the GPUs accounts for the activity on the internal IP cores in some proprietary way and reports it back to the system. (that's similar on how most GPUs work since they're all proprietary IP cores).

[Eigenvektor]

23 minutes ago, Edgar R. Zakarian said:

like number of active transistors?

Pretty sure that would be quite impractical (and complicated). You'd need a ton of additional circuitry to measure and count which transistors are active, which would then need additional space, require additional power and produce more heat.

 

The CPU load is determined by the OS and how it does this depends on the particular OS you're using. If you wanted to go really in-depth, you could probably have a look at how the Linux kernel does it (provided you can read and understand its code).

 

Here's an answer that is a bit more in depth: https://stackoverflow.com/a/3748172

[Moonzy]

i would assume it's the ratio of your cpu having work queued vs having nothing queued for the cycle across a fixed amount of time

same goes for gpu -shrug-

 

as for HT, a cpu can process more than 1 type of instruction at once so HT can queue different types of workload for a core to process at once.