Keryx Miner Supr v0.16.3 — per-card H16 solver tuning
This release adds faster NVIDIA and AMD H16 walk implementations and chooses the
fastest exact implementation on each card. Existing kernels remain available.
- NVIDIA: reuse tile storage and pack Merkle-parent work. Warmed per-card timing
selects the new CUDA kernel only when it wins by at least 0.5%. - AMD: cooperative INT8 matrices, shared address resolution, and subgroup dot
reductions. Per-card/model tuning selects the kernel and a 1,024/4,096/8,192
nonce batch while keeping the measured dispatch within 20 ms. - AMD mining, inference and audit replay continue on the same GPU. Both the
regular and pre-AVX Vulkan engines are included; use the complete package. - H16 activation, consensus rules and model weights are unchanged.
Matched H16 tier-0 development measurements against v0.16.2:
| GPU | Power limit | v0.16.2 MH/s | New solver MH/s | Change |
|---|---|---|---|---|
| RTX 5090 | 420 W | 5.2208 | 5.4771 | +4.91% |
| RTX 5090 | 600 W | 6.2054 | 6.2035 | Neutral |
| RTX 5080 | 360 W | 3.3953 | 3.3985 | Neutral |
| RTX 3070 | 240 W | 1.5696 | 1.5693 | Neutral |
| RX 7900 XTX | 272 W | 0.6149 | 1.0945 | +78.00% |
| RX 7600 XT | 155 W | 0.2874 | 0.4456 | +55.02% |
| Radeon Instinct MI50 | 150 W | 0.2721 | 0.4421 | +62.48% |
These are matched development measurements of the solver included here, not a
fresh benchmark of every packaged platform. Full-power 5090, 5080 and 3070 kept
the previous two-nonce kernel. See the benchmark methodology, clocks, measured
power and CPU use.
Qualification includes CPU/GPU exactness, 400 captured mining proofs checked by
the unmodified node 1.6.7 consensus verifier, and full H16 inference, attention,
custody, restart and mining-resume checks on RTX 5090 and all three AMD models.
Windows GPU execution and higher-tier performance were not measured. macOS
remains an experimental Metal package with its existing H15/H16 limitations.
Packages: Linux NVIDIA modern and legacy, Linux AMD, HiveOS, mmpOS, SMOS,
Windows NVIDIA and AMD, macOS arm64, and Docker
ocminersupr/keryx-miner-supr:0.16.3 (latest). Upstream licenses and earlier grants are preserved.
Original contributions in this revision use the Keryx Miner Supr
Source-Sharing License 1.0: no closed-source incorporation; source-sharing
derivatives must provide corresponding source, credits and a repository
backlink. Running the original miner for profit is permitted. Previously
granted rights remain valid; see the license scope.
Final Linux package checks passed H16 mining/inference/custody/replay on RTX 5090
and RX 7900 XTX. The post-H16 live pool test accepted 75 shares with zero rejects.
All 270 host tests and 33 Ubuntu 20.04/22.04/24.04 startup checks passed.