Question
Is there any officially supported way to set the board power cap/PPT of an
AMD Instinct MI210X while its PF is managed by GIM and its single VF is passed
through to a VM?
I would like to test 275 W, 250 W, and 225 W limits instead of the default
300 W.
Environment
- GPU: AMD Instinct MI210X, PF device 1002:740f, one VF 1002:7410
- VBIOS/IFWI: 113-D67301V-075
- GIM: 9.1.0.K, staging commit
18fb8f253533a1a6425d2b361d53980d7e3e9e7e
- Host: Proxmox VE 8.4, kernel 6.8.12-39-pve
- Guest: Ubuntu 22.04, kernel 5.15.0-94
- Guest driver: amdgpu 6.16.13, ROCm 7.2.4
- Guest AMD SMI: 26.2.2, Platform: Linux Guest
I understand that this PVE host is outside the officially validated host OS
matrix. My question is about the MI200 implementation and supported management
interface in GIM 9.1.0.K; the observations below also come directly from that
release's API and source.
Observed behavior
Inside the guest, as root:
sudo amd-smi set --gpu 0 --power-cap ppt0 250
AMDSMI_STATUS_NO_PERM
The CLI reports a PPT0 maximum of 300 W, but the operation is not permitted in
the guest, which is consistent with the AMD SMI documentation.
On the host PF, using libamdsmi.so built from the same GIM source:
current power cap: 300 W
default: 0 W
minimum: UINT64_MAX
maximum: UINT64_MAX
amdsmi_set_power_cap(..., 275 W): AMDSMI_STATUS_NOT_SUPPORTED
The power cap remains 300 W after the call.
In the current staging source, mi200_gpumon_funcs registers
.get_gpu_power_capacity but does not register .set_gpu_power_capacity.
The MI300 implementation registers both callbacks. I also noticed that the
MI200 SMU header defines SetPptLimit, but I could not find a supported MI200
setter wired to the host AMD SMI API.
Clarification requested
- Is per-GPU power-cap control intentionally unsupported for MI210X under GIM,
or is the host setter not implemented yet?
- Is there another officially supported host-side interface, GIM parameter,
or out-of-band management method for setting MI210X PPT under SR-IOV?
- If a power cap is not supported, is an official GFX clock limit supported
for this configuration as an alternative efficiency control?
- Is MI200 power-cap setter support planned for a future GIM release?
Thank you.
Question
Is there any officially supported way to set the board power cap/PPT of an
AMD Instinct MI210X while its PF is managed by GIM and its single VF is passed
through to a VM?
I would like to test 275 W, 250 W, and 225 W limits instead of the default
300 W.
Environment
18fb8f253533a1a6425d2b361d53980d7e3e9e7eI understand that this PVE host is outside the officially validated host OS
matrix. My question is about the MI200 implementation and supported management
interface in GIM 9.1.0.K; the observations below also come directly from that
release's API and source.
Observed behavior
Inside the guest, as root:
The CLI reports a PPT0 maximum of 300 W, but the operation is not permitted in
the guest, which is consistent with the AMD SMI documentation.
On the host PF, using
libamdsmi.sobuilt from the same GIM source:The power cap remains 300 W after the call.
In the current staging source,
mi200_gpumon_funcsregisters.get_gpu_power_capacitybut does not register.set_gpu_power_capacity.The MI300 implementation registers both callbacks. I also noticed that the
MI200 SMU header defines
SetPptLimit, but I could not find a supported MI200setter wired to the host AMD SMI API.
Clarification requested
or is the host setter not implemented yet?
or out-of-band management method for setting MI210X PPT under SR-IOV?
for this configuration as an alternative efficiency control?
Thank you.