Windows walk hashrate fix. Windows NVIDIA rigs were mining the H6 walk at ~1.2 kh/s vs ~6.4 for other miners ("limited on power"). This release fixes that — no change to Linux/HiveOS.
What was wrong
The walk kernel shipped as compute_75 PTX and relied on the GPU driver JIT-compiling it to sm_120 (Blackwell) at load. On Linux that JIT is good (~7.5 kh/s on a 5080). On Windows the driver's JIT of the same PTX to sm_120 is catastrophically bad (~1.2 kh/s).
The fix
Windows builds now ship native sm_120 SASS for the walk (a prebuilt multi-arch fatbin: sm_75/80/86/89/90/120 + PTX fallback), so there is no driver JIT on Blackwell — the walk runs at full speed (~6.9 kh/s on a 5080-class card, matching/beating other miners). Validated byte-exact (pool-accepted shares).
Linux/HiveOS/MMPOS is unchanged — it keeps the PTX path, whose JIT is ~9% faster than offline native SASS, so there is zero fleet regression. Only the Windows-NVIDIA build changed.
Who should update
- Windows NVIDIA (Blackwell: 5070 Ti / 5080 / 5090) — grab
keryx-miner-supr-windows-nvidia-pom.zip; expect a ~5× walk hashrate jump. - Linux/HiveOS — optional (walk identical to v0.10.7; also carries the v0.10.7 inference-serving + VRAM-fit fixes).
Builds: modern (CUDA 12.9, sm_75–120) + legacy (CUDA 12.4, sm_70+), HiveOS/MMPOS/Linux + Windows + AMD.