CPU/GPU Usage Between NVIDIA & AMD Linux Drivers

[phoronix]

Phoronix: CPU/GPU Usage Between NVIDIA & AMD Linux Drivers

Following the 4K AMD/NVIDIA High-End GPU Comparison On SteamOS Linux and 22-Way Comparison Of NVIDIA/AMD Graphics Cards On SteamOS For Steam Linux Gaming articles, a few Phoronix readers were inquiring about the CPU and GPU utilization metrics during testing...

http://www.phoronix.com/scan.php?pag...-CPU-GPU-Usage

[atomsymbol]

Phoronix: CPU/GPU Usage Between NVIDIA & AMD Linux Drivers

Following the 4K AMD/NVIDIA High-End GPU Comparison On SteamOS Linux and 22-Way Comparison Of NVIDIA/AMD Graphics Cards On SteamOS For Steam Linux Gaming articles, a few Phoronix readers were inquiring about the CPU and GPU utilization metrics during testing...

http://www.phoronix.com/scan.php?pag...-CPU-GPU-Usage

Executing "aticonfig --odgc --odgt" can be used to print GPU load and temperature. PTS 6.0.0m3 seems not to be using this.

[magika]

980 Ti more CPU usage, but also highest fps so that CPU usage is expected. Probably normalized per fps will be easier to see.
But again, I played redux version of first metro game for not more than a hour (because its incredibly retarded & boring), but that game is probably "opengl port done right".

[curaga]

Before asking for open-source usage numbers, with the Radeon/Nouveau DRM drivers there also doesn't seem to be a way to read the global GPU utilization as a percent.

Radeontop, the support is already in PTS. It should even report the percentage correctly when running on the blob.

[gamerk2]

Question: What are you measuring: The GPU driver load, or the total game load? Two different things. You need to isolate the GPU threads if you really want to see what's going on in driver land.

[efikkan]

Hi Michael,

Are you trying to measure the CPU "overhead" for different GPUs/drivers? Because you wouldn't be able to measure it this way. Most performance hungry games will fully utilize their thread with GPU context, regardless of how efficient the driver is. You would not be able to determine how much of this load is spent waiting for the driver/GPU and how much is actually processing, you'll to access the driver and/or game engine to determine that. Most games will have their GPU thread at 100% load to minimize latency and scheduling problems. Most people mistake this for a game being CPU limited. Games could reduce this "load" by adding a sleep between frames, but that may introduce micro stuttering.

[haagch]

Are you trying to measure the CPU "overhead" for different GPUs/drivers?

Finding abnormal bottlenecks. The GPU usage on starting is quite erratic, perhaps it's still loading or compiling shaders, but then it looks consistently high, so we can conclude that on nvidia metro redux uses the gpus very well. (capping percentage graphs at 100% would look nicer by the way)

From last time I run the benchmark I predict that on radeonsi in contrast we see very low GPU usage, suggesting there are opportunities in the driver to make it run much faster. Maybe I'll try it later to get some data.

edit: And on open source radeon driver:

sudo radeontop -d - | cut -f 3 -d " "
old
-,
0.00%,
0.00%,
0.83%,
0.00%