Keryx Miner Supr v0.16.5 — AMD RX 5700 inference fix
Recommended update for AMD rigs, required for RX 5700 / 5600 (RDNA1). NVIDIA
rigs gain nothing new; their mining and inference are unchanged from v0.16.4.
- Fix: GPU inference on AMD cards without hardware int8 dot products. On
RX 5700 / 5600 (Navi10) the Linux AMD Vulkan driver (RADV) offers int8 dot
products but reports them as not hardware-accelerated. The Vulkan inference
engine then dropped its int8 path, so the startup self-test failed with
did not take the q8_1 mat-vec path, the model was withdrawn and every card
of the rig stayed at 0 H/s ("preparing"). The engine now keeps the int8 path
on these cards. Integer results are exact, so the inference traces are the
same as on any other card. Cards with hardware int8 dot (RX 6000/7000/9000,
MI50 and newer) behave exactly as before. - HiveOS AMD: the OpenCL preflight now counts GPUs correctly with ROCm's
clinfo. Before, it could print a false
no OpenCL devices visiblewarning although the miner found every card. - The OpenCL message
Out-of-order queue unsupported, falling back to in-order
is now an info line. It was always harmless. - Mining kernels, H16 rules and model weights are unchanged from v0.16.4.
Tested with the release packages on an RX 7900 XTX, with the engine told to
treat the card as non-accelerated like an RX 5700: the traced self-test passed
and the card mined to the pool with accepted shares and 0 rejected. Without the
fix the same test reproduced the reported failure. We have no RX 5700 to test on;
the miner's startup self-test checks every card before it serves or mines a tier.
Windows GPU execution was not tested.
Packages: Linux NVIDIA modern and legacy, Linux AMD, HiveOS, mmpOS, SMOS,
Windows NVIDIA and AMD, macOS arm64, and Docker
ocminersupr/keryx-miner-supr:0.16.5 (latest).
Licensing is unchanged from v0.16.3: original contributions use the Keryx
Miner Supr Source-Sharing License 1.0 (see the license
scope);
third-party licenses and earlier grants are preserved.