AMD card support - looking for per VF point voltage offset API

[jpietek]

So far some basic research done with LLMs:

https://jpietek.github.io/PenguinBurner/amd-rdna-undervolting/

So far the biggest problem it seems is the limited amdgpu+kernel capabilities with per point vf curve getters and setters.

But we could do some poor man's static uv and simulate the Nvidia behaviour just with global clock and voltage offsets.

More interesting though would be a possiblity to hack more fine grained n-zones offsets to shape the curve better, any insights/hints on that are highly appreciated! :)

[jpietek]

Research update (2026-08-11): AMD RDNA undervolting on Linux — what exists, what we can hack, and what we can visualize

The report covers the exact stock Linux 7.1.8 global vo surface, relevant 7.2/development-tree changes, the proposed SMU13/14 six-boundary setter and getter (five PWL zones / six boundary-offset nodes), available AMD telemetry, measured-response visualization, and a capability-based AMD/NVIDIA adapter. RDNA4 is the first experimental target, RDNA3 follows, and RDNA2 remains the simpler capability-probed global-offset path.

This is source-level research; RX 9060/9070 hardware qualification and proof that PMFW accepts unequal boundary values are still pending. Published via PR #44.