Express memory offset in reported-clock domain, convert 2x at NVML boundary

On GDDR7/Blackwell (measured: RTX 5090, driver 615.78.08) the memory
V/F offset NVML APIs operate in a domain 2x the memory clock NVML
reports. Applying +3000 moved the under-load clock by only +1500 MHz;
the driver range -2000..+6000 maps to NvAPI's -1000..+3000.

- limits.py: multiply by _MEM_OFFSET_DOMAIN_FACTOR (2) on set, divide
  on get/readback/range for the MEM domain only; GPC untouched.
  API/CLI/profiles/UI now all speak the reported-clock domain
  (what nvidia-smi shows), so a +3000 slider value reaches the
  driver's full +6000 headroom (old UI half-used it).
- limits.py: call nvmlDeviceGet/SetClockOffsets via the raw function
  pointer for the real driver return code - nvidia-ml-py wrappers
  discard it and always report success, which would have masked
  erroring drivers and skipped the deprecated-API fallbacks.
- PerformancePanel: fallback range -1000..+3000, remove stale
  Math.min(max, 3000) cap that clamped the new max, add domain tooltip.
- Docs: document the domain conversion (API.md, Usage-Guide.md).
This commit is contained in:
ARIA committed 2026-11-12 13:43:17 +01:00
1 parent 2e8e580d79
commit a76685a10b
4 files changed
+134 -58

No files matched your search

+4
View File
@@ -98,6 +98,10 @@ Notes:
Clients cannot raise it per request.
- **Units** — curve deltas are in **kHz**; clock offsets and memory offsets
are in **MHz**; voltages in µV/mV; power in watts.
The memory offset (`mem_offset_mhz`) is in the *reported-clock* domain —
the value shown by NVML memory clock / `nvidia-smi`. On GDDR7/Blackwell
the driver internally applies 2x that amount (see
`_MEM_OFFSET_DOMAIN_FACTOR` in `nvcurve/hal/limits.py`).
- **Side effects** — write endpoints (`/api/curve/*`, `/api/limits`,
`/api/profiles/*/apply`, `/api/snapshot/restore`) change hardware state.
With `auto_snapshot: true` (default) a ClockBoostTable snapshot is saved
+2
View File
@@ -88,6 +88,8 @@ Data is streamed via WebSocket from the backend at a configurable poll interval
The Performance panel controls the board power limit and the memory clock offset. The power limit slider is bounded by the GPU's VBIOS minimum and maximum (shown at the slider ends); changes are applied on **Apply** and reset to the hardware default on **Reset**.
Memory clock offsets are expressed in the *reported-clock* domain — the value `nvidia-smi` shows for memory clock. On GDDR7/Blackwell the driver operates in an internal domain that is 2x that: a +3000 offset raises the reported (nvidia-smi) clock by +3000 MHz, while the driver applies +6000 internally. API, CLI and profiles all use the reported domain.
#### Experimental NVIDIA power control
On compatible drivers, an **Experimental NVIDIA power control** checkbox appears in the Performance panel. Enabling it switches power-limit application from the standard NVML call to an undocumented driver (RM) interface, which **permits caps below the VBIOS minimum, down to 30 W**. The native maximum still applies.