Power Requirements¶
Power feed for a CMP 170HX. The watt figures are the published TDP and the software cap. Meter your own wall draw under load.
Sources: Full Specifications, Consensus-Protocol/cmp170hx.
Envelope¶
| Item | Value |
|---|---|
| Default TDP | 250 W |
| Software max power limit | ~300 W |
| Idle (typical) | ~30–40 W |
| Stock connector | 1× 8-pin CPU/EPS-style via adapter |
| Slot power | Do not rely on the slot alone for load |
EPS-style plug
The card expects the mining/server adapter path: an 8-pin CPU/EPS-style connector. Use the shipped adapter or a verified equivalent and check seating twice before first power-on.
Sizing the PSU¶
Plan continuous budget as:
Practical floors when shopping used PSUs:
| Host class | Suggested 12 V headroom |
|---|---|
| Single 170HX + modest CPU | 550 W quality unit minimum comfort |
| Single 170HX + workstation CPU + dual NVMe | 650–750 W |
| Dual 170HX | Size for 2 × 300 W GPU ceiling plus host, then add margin |
Idle is friendly (~30–40 W on the GPU). Sustained unlocked FP32 / LLM decode is where the 250–300 W band shows up. Meter the wall once during a long generate job and write the number next to your Performance entry.
Software limit vs heatsink
Raising the power limit toward 300 W without proving cooling first is how you meet the thermal runaway warning on Air Cooling / Water Cooling. Leave the limit at 250 W until the card holds a steady temperature under load.
Cabling checklist¶
- EPS/CPU 8-pin fully latched on the adapter
- Adapter fully latched on the card
- No SATA-to-PCIe cheater cables anywhere in the GPU path
- Cable gauge appropriate for ~25 A class 12 V draw on that branch
- Strain relief so the stiff mining cable cannot lever the PCB (especially after waterblock installs)
Multi-GPU notes¶
Each additional 170HX adds another ~250–300 W planning chunk and another EPS feed. P2P / BAR1 experiments (see forks such as bayley/cmpunlocker) still need the same electrical budget. Gen1/Gen2 x4 links will dominate communication long before the PSU does if you skipped the capacitor mod.
Failure modes¶
| Symptom | Likely cause | First check |
|---|---|---|
| Instant shutdown under load | OCP / weak 12 V rail | Meter PSU rating; reseat EPS |
| Card disappears after resume | Brown-out or cable creep | Reseat adapter; check dmesg |
| Connector warm to the touch | High resistance crimp or partial seat | Power off; inspect pins |
| Idle power >> 40 W | Stuck clocks / failed fan curve on host / measuring whole system | Isolate GPU via nvidia-smi -q -d POWER |