Keryx Miner Supr v0.16.8 — faster AMD H16 mining
- AMD Vulkan walk: each step now finds its 1 KiB tile through a small bucket
table (one load and a short scan) instead of a binary search over the
tensor table. At start-up every card cross-checks the new lookup against
the previous search on its own model and keeps the search if they differ. - AMD Vulkan walk: one exact modulo per step (context pad or weights) instead
of computing both. - Linux AMD: the per-card dispatch tuner also tries 16384 and 32768 nonces per
launch, still bounded to 20 ms per launch. Windows keeps its previous bound.
Measured against v0.16.7 on the same card, power limit and tier-0 model,
matched before/after/after/before runs of 120 s (MH/s):
| GPU | Power limit | v0.16.7 | v0.16.8 | Change |
|---|---|---|---|---|
| RX 7900 XTX | 350 W | 1.225 | 1.502 | +22.6% |
| RX 7600 XT | 155 W | 0.442 | 0.506 | +14.6% |
| Instinct MI50 | 150 W | 0.439 | 0.461 | +5.0% |
The MI50 runs were shortened to 65-105 s by the test's temperature limit.
Larger tiers gain similarly on the development build (RX 7900 XTX tiers 1-3,
RX 7600 XT tiers 1-2). Results depend on cooling, power limit and driver.
KERYX_VK_V5_BUCKET=0 restores the table search. The RDNA1 descriptor path,
NVIDIA kernels, H16 consensus, model weights, inference, custody and per-card
pause/proof requirements are unchanged; NVIDIA native engines are identical to
v0.16.7. macOS retains its experimental Metal/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:0.16.8 (latest). Licensing remains
Source-Sharing License 1.0; upstream licenses and previous grants are preserved.
Validation: 285 host/integration tests, 33 Linux package startup checks on Ubuntu
20.04/22.04/24.04 and 14 archive audits. Final Linux AMD package on RX 7900 XTX:
CPU-reference walk tests, isolated H16 traced inference with all audit opening
phases, custody upload/restart/replay and mining resumption (also with the pre-AVX
engine in RDNA1 compatibility mode, at a reduced power limit), the context-refusal
reply, and a live TLS pool check with 9 accepted and 0 rejected shares. RX 7600 XT
and Instinct MI50: CPU-reference tests on both cards and two-card inference plus
mining in both serving orders. On mixed rigs an MI50 can intermittently lose its
Vulkan device at start-up; a control run shows the same with v0.16.7, so this
release does not change it (restart the miner). Windows NVIDIA/AMD and macOS
packages were built and tested by CI; the Windows CUDA inference engines are
identical to v0.16.7. These are functional checks plus the measurements above,
not on-chain AI request inclusion claims.
Physical RDNA1/RX 5700, Windows GPU execution and ten-card
hardware remain unqualified. Existing benchmark tables retain their original
version/workload labels.