Supported Gpu Matrix
| Vendor | Family | Models | Status |
|---|---|---|---|
| NVIDIA | A100 | SXM4, PCIe, NVL | Full (NVML+DCGM) |
| NVIDIA | L40 / L40S | PCIe | Full (NVML+DCGM) |
| NVIDIA | L4 | PCIe | Full (NVML+DCGM) |
| NVIDIA | T4 | PCIe | Full (NVML) |
| NVIDIA | V100 | SXM2, PCIe | Partial (NVML; no DCGM MIG iteration) |
| NVIDIA | A40 | PCIe | Full (NVML) |
| NVIDIA | H200 | SXM | Full (NVML+DCGM) |
| NVIDIA | B100 / B200 | SXM | Best effort (NVML; DCGM 3.3+) |
| NVIDIA | GB200 NVL72 | NVL | Best effort (NVML; FM REST) |
| AMD | MI250X | OAM | Best effort (rocm-smi) |
| AMD | MI300X | OAM | Best effort (rocm-smi) |
| Intel | Gaudi 2 / 3 | PCIe | Best effort (hl-smi) |
MIG Slices
Per-instance iteration is supported for A100 and H100. Each MIG slice gets its own GPU-UUID-suffixed stream.
Multi-Instance GPU (MIG)
- A100: 1g.5gb, 2g.10gb, 3g.20gb, 4g.20gb, 7g.40gb.
- H100: 1g.10gb, 1g.20gb, 2g.20gb, 3g.40gb, 4g.40gb, 7g.80gb.
Each slice is queried independently; metrics are tagged with the parent GPU's UUID plus the instance id.
Unsupported
- Driver 470 or older (no NVML v2/v3 symbols).
- GPUs without ECC (some consumer SKUs).
- Cross-vendor MIG (MI300X virtual CDUs are not yet mapped).
Repro
docs/operations/gpu-matrix-probe.sh runs on a target host and emits a JSON manifest of what is in scope.