Per-card batch control, and a mode for inference-first rigs
--intensity — set the grind batch by hand
Works the way cgminer/sgminer users expect: batch = 2^intensity, one value per card.
--intensity 18,18,16 # GPU0/GPU1 = 262144 nonces, GPU2 = 65536Positions follow the cards this process mines, so with --cuda-device 2,3 the first value is GPU2.
An empty slot (18,,16) leaves that card on autotune, and a card you list is not benchmarked at all.
Higher is not automatically better — on an RTX 5080 the measured optimum is 64512, and forcing
intensity 17 gives 2.95 MH/s against the autotuned 2.99.
The batch is capped to what the card's free VRAM can back. This is a safety limit, not a
preference: on an 8 GB RTX 3070, --intensity 18 puts 512 MiB of offset buffers on a card already
holding a 6.4 GB model, which collapsed it from 1.35 MH/s to 26 kH/s and left the CUDA context
poisoned — thousands of illegal-access errors, with inference dead on that GPU afterwards. An
unattainable intensity is now clamped with a warning and the card keeps running normally.
--only-inference — serve requests, barely mine
For rigs that are here for OPoI rewards rather than PoW. The walk drops to the smallest batch with an
idle pause between launches, and stops completely while a request is being served, so the card
answers at full speed and then goes back to a trickle.
Measured on an RTX 5080, same card and model tier:
| mode | hashrate | power | temp |
|---|---|---|---|
| normal | 2.99 MH/s | 400 W | 70 °C |
--only-inference
| 32 kH/s | 58 W | 41 °C |
Roughly 14% of the power, with the card cool and free to serve. Tune the idle share with
KERYX_ONLY_INFERENCE_DUTY_MS (default 250). The no-share wedge supervisor is disabled in this mode,
since a rig that is deliberately barely hashing would otherwise be restarted every 10 minutes.
The grind batch now starts high
A card that has not benchmarked itself yet runs 768 nonces/SM instead of 384. Every card measured
from ~66 SMs up prefers the larger batch — by up to ~4% — and only a small 46-SM card prefers less, so
starting at the top of the sweep means most cards are already at their best before the autotune has
said anything, and the tuner's job is to walk a card down rather than creep up to it.
Upgrading is safe for existing rigs: with no --intensity and no --only-inference, behaviour is
unchanged apart from the higher starting batch and the VRAM cap.