Add AMD GPU collector

[Dim-P]

Summary

This PR adds a collector that monitors all the AMD GPU cards on the system.

Fixes #12616.

The collector provides the following metrics:

• Utilization
  • GPU utilization
  • GPU memory utilization
• Clock frequencies:
  • GPU clock frequency
  • GPU memory clock frequency
• Memory usage:
  • VRAM memory usage percentage
  • VRAM memory usage
  • visible VRAM memory usage percentage
  • visible VRAM memory usage
  • GTT memory usage percentage
  • GTT memory usage

The following can be provided by Sensors, so they were not implemented:

• Power consumption
• Temperature
• Fan speed
• Voltage

Out of scope of this PR, but could potentially be implemented in the future (needs further research):

• PCIe RX/TX bandwidth metrics
• Overclocking metrics
• Usage and consumption per process (similar to nvtop)
• Power state metrics

Test Plan

• Run this PR on a system without an AMD GPU and check that this branch behaves as master.
• Run this PR on a system with an AMD GPU and check that all the above metrics are displayed. Check also that the "marketing name" for each GPU is correct (can be checked at the product_name label).

For users: How does this change affect me?

You will be able to monitor basic metrics for AMD GPUs. It does not affect any instances that do not have an AMD GPU (see the nvidia-smi collector for NVIDIA GPUs).

[Dim-P]

@ilyam8 Removed configuration as discussed. Please let me know if chart IDs and contexts make sense as they are, I think they do now. I tested with multiple GPU cards and aggregated metrics seem to work correctly.

[MrZammler]

Hi Dimitri! Nice!

So, I tested this on my laptop, which has an AMD 4500U CPU, with integrated GPU.

The card is on this list.

I've added {0x1636, 0xC3, "AMD 4500U"}, in your list, since it wasn't included. So maybe there are some more cards that could be added, but that can be iterated.

What I've found next, is that still it wouldn't display any charts for it. The problem was that in my case util_mem does not exist. Commenting out the part for checking util_mem, did allow other charts to appear (gpu_utilization, gpu_clk_frequency and gpu_mem_clk_frequency).

So a suggestion would be that if in case something is missing, don't give up on the whole card?

Also, if possible to capitalize the amdgpu entry in the list on the dashboard (e.g. AMDGpu) and an icon would be cool as well :-)

[Dim-P]

Hi Dimitri! Nice!

So, I tested this on my laptop, which has an AMD 4500U CPU, with integrated GPU.

The card is on this list.

I've added {0x1636, 0xC3, "AMD 4500U"}, in your list, since it wasn't included. So maybe there are some more cards that could be added, but that can be iterated.

What I've found next, is that still it wouldn't display any charts for it. The problem was that in my case util_mem does not exist. Commenting out the part for checking util_mem, did allow other charts to appear (gpu_utilization, gpu_clk_frequency and gpu_mem_clk_frequency).

So a suggestion would be that if in case something is missing, don't give up on the whole card?

Also, if possible to capitalize the amdgpu entry in the list on the dashboard (e.g. AMDGpu) and an icon would be cool as well :-)

Thanks, good points.

About additional GPU marketing names, I believe these can be added later, as it's a never-ending piece of work. I did add some RX 7xxx though that were missing from the original list.

I changed how chart updates are performed and now a linked list is used, so any metrics (and charts) that cannot be monitored any more can easily be removed from the list. This seems to be working fine but please test on your system again.

About amdgpu, it is capitalized if you check the first version of the dashboard (localhost:19999/v1), but I believe for v2, this (and the icon) is handled by @netdata/cloud-fe .

I also checked @Ferroin 's concerns about introducing unnecessary GPU usage due to /sys/class/drm calls, but I didn't notice any difference on my setup.

[MrZammler]

Seems to work fine here! Just needs more cards, we can add later!

[thiagoftsm]

Please, do not forget to create an analogous PR to also bring this description for new dashboard.

[Dim-P]

Thanks, will do, just need the information of what needs updating from @netdata/cloud-fe as mentioned in my comment above.

[ilyam8]

Minor: we shouldn't update this file in netdata/netdata but in https://github.com/netdata/dashboard. If there is a new version of the old dashboard, this file will be overwritten.

[thiagoftsm]

I do not have hardware to test, bu considering that other teammate already tested it, and possible changes will come in other PRs, LGTM!