ocminer/keryx-miner-supr v0.6.6.0

keryx-miner-supr v0.6.6.0

on GitHub

keryx-miner-supr v0.6.6.0

Minor release folding in the H100 / Proof-of-Model tuning work, plus a new open GPU benchmark table.

What's new

• PoM throughput: POM_BATCH raised to 2²² (≈ +2.5 % on high-bandwidth cards, measured live on H100).
• H100 / datacenter tuning: microbench harness (tools/h100/) proving the walk is at the HBM memory ceiling; capping an H100 to ~400 W keeps full hashrate for ~30 % less power.
• BENCHMARKS.md: a new, openly PR-editable per-card table — card, architecture, generation, VRAM, mem/core clock, mining power, PoM hashrate and efficiency — from our own fleet + H100 measurements, with tuning notes and the full H100 power sweep. Miners are welcome to submit corrections and add their cards.
• Windows: the Windows zips now ship a ready-to-edit start.bat (replace REPLACE_ME with your wallet + .rigName) that keeps the console open on exit/crash.

Downloads — pick your line

Three NVIDIA build lines, each carrying its own bundled CUDA runtime (driver floor):

• modern — CUDA 12.9, driver R575+, native Blackwell (sm_120). For 50-series / newest drivers.
• legacy — CUDA 12.4, driver R550+. For Turing/Ampere/Ada on older drivers (forward-JITs to Blackwell).
• pascal — CUDA 12.4, sm_60 PTX. For GTX 10-series / P100 (run --cpu-inference --light).

Each line ships in 4 formats: HiveOS, SMOS, mmpOS, and generic Linux x86_64. Verify with the matching SHA256SUMS-<line>.txt. Windows zips (AMD, NVIDIA-PoM, keryxcuda) are attached automatically by CI.

Cargo/pool user-agent version stays 0.6.5; the release label (CUSTOM_VERSION) is 0.6.6.0.