keryx-miner-supr 0.14.2

Performance and stability release for the H14 lineup. Recommended for every rig, especially Ampere data-center cards (CMP 170HX, A100-class), mixed rigs where one card has less VRAM than the others, and AMD rigs with 16 GB cards.

Faster

• New PoM walk backend sidecar-chase+addr-tensor-core (+5.5–5.8 % on CMP 170HX). Profiling showed the walk kernel is limited by integer instructions, not memory bandwidth, and about an eighth of each step went into finding which model segment a tile lives in (a four-step dependent lookup). The miner now builds one address per tile at load time and jumps straight to it; tiles that cross a segment boundary keep the old exact path, so the bytes fetched are identical. It is an autotune candidate like the other walk backends: each card measures it at startup and only uses it when it is faster and its results match the reference. Measured on a CMP 170HX with Kimi-Linear-48B, live on the pool for 20 minutes each: 0.14.1 2.43 Mh/s → 0.14.2 2.58 Mh/s (+5.9 %, +7.9 % per MHz of core clock), shares accepted, 0 rejected. Other cards keep whatever the autotune picks; RTX 50-series skip the sidecar backends as before.

Fixed

• One card running out of memory no longer stops the whole rig. When a single card could not build its PoM walk (walk_build_oom), the model was withdrawn for the entire rig, so healthy cards on the same model stopped mining too, and the next self-test re-announced it — the rig flapped every ~30 s. Now only the card that ran out of memory is moved to the next smaller downloaded tier; the model stays announced while another card still mines it.
• AMD: the card reserved for inference is no longer shown as STALLED. When the inference model and the PoM walk cannot share one card (for example Gemma-4-12B on 16 GB RX 6800 / 6800 XT cards), the miner deliberately dedicates one card to answering inference requests and mines on the others — that card was always meant to show 0 H/s, but the dashboard marked it with a red STALLED alert. It is now shown as an inference card (INF). A card that really stops hashing is still shown as STALLED. Verified on a rig with three 16 GB AMD cards.

Which download

No consensus, protocol or model changes; 0.14.x miners can run side by side. H14 gate (DAA 121,985,000) handling is unchanged from 0.14.0. The first start after the upgrade re-runs the walk autotune (about a minute).