Skip to content

RIG_PROFILES.md

Latest commit

 

History

History
53 lines (37 loc) · 2.13 KB

RIG_PROFILES.md

File metadata and controls

53 lines (37 loc) · 2.13 KB
Raw

Rig profiles (GPU pool tuning)

Built-in profiles tune pool worker env (hackme.env) — not BIOS clocks directly.

API

Method Path Auth
GET /api/hardware/rig-profiles —
GET /api/hardware/rig-profiles/detect — (GPU names, vendor, suggested_backend, hybrid notes)
POST /api/hardware/rig-profiles/apply admin

Apply merges profile keys into hackme.env, then stop/start pool worker from the Mining tab.

Stage 1 — auto-detect (no manual NVIDIA/AMD flag)

On node startup, if hackme.env has no rig tune keys, HackMe:

  1. Probes GPUs via sysfs (/sys/class/drm), lspci, NVIDIA /proc, AMDGPU telemetry, WMI (Windows).
  2. Picks backend: NVIDIA + CUDA binary → cuda; AMD / Intel → opencl; none → cpu.
  3. Merges matching rig profile into hackme.env (disable with HACKME_RIG_PROFILE_AUTO=0).

Built-in profiles include: RX 580 (2048SP), Intel Arc, AMD RDNA3, RTX 30/40/50.

curl -sS http://127.0.0.1:8080/api/hardware/rig-profiles/detect | jq .

AMD RX 580 2048SP (Yaroslav / Chinese refresh)

Profile id: amd_rx580_2048sp

Env Value Why
HACKME_GPU_BACKEND opencl (Linux) / auto (Windows until OpenCL binary ships) Polaris OpenCL
HACKME_WORKER_BATCH_SIZE 1048576 Smaller batches for 2048SP memory
GPU_CHUNK 524288 Shorter GPU search slices
SEARCH_TIMEOUT_MS 4500 Avoid timeout on slow kernels
HACKME_WORKER_CLAIM_COOLDOWN_MS 150 Fewer 429s on sub-1 GH/s rigs
HACKME_GPU_TEMP_PAUSE_C 78 Thermal hold
HACKME_GPU_HASHRATE_FLOOR_GHS (optional) Minimum GH/s floor only if kernel timing fails — do not pin pool difficulty

Manual OC (vendor tools): core 1150–1200 MHz, mem 2000–2100 MHz effective, power −5…−8%.
After stable OC, switch to amd_rx580_2048sp_turbo in the dashboard.

Verify live metrics

bash scripts/ops/verify_pool_dashboard.sh
curl -sS http://127.0.0.1:8080/api/hardware/rig-profiles/detect | jq .

Mining tab shows Difficulty (M), GH/s, and rig profile status updating on poll.