Upgrade before mainnet DAA 147,000,000 (testnet 180,000) for compatibility with keryx-node 1.6.7. The miner continues normally below the gate and automatically switches to H16 at activation, freeing 200 MiB of context-pad VRAM per tier. Reorgs and old jobs retain the rules for their own DAA.
Solo mining supports the new awaited-request RPC, scheduling across rigs sharing an address, transient reconnect recovery and persisted audit extensions. --testnet now selects the full testnet gate table. Inference stays on the selected GPU and fits its context to available memory.
This release includes the CPU/IPFS activity fix and exact CUDA solver/cache improvements. Matched H15 development measurements: RTX 5090 +3.6–3.8%, RTX 5080 +0.6%, RTX 5070 Ti +2.0%, RTX 3070 +4.5%; later cache work added +5.15% on CMP 170HX and +0.70% on RTX 3070 relative to the preceding candidate. Models, power settings and qualification limits are documented in the README. These are development comparisons with our own previous candidates, not H16 network benchmarks.
Validation: 15 final GPU proofs passed the unmodified node v1.6.7 consensus verifier before, at and after the official mainnet gate, including rollback and recrossing. Wrong-era pads, tampering and one-below targets were rejected. Both official network boundaries passed native GPU/CPU checks. H16 traced inference, independent proof/attention checks, custody fetch after restart and mining resumption passed on RTX 5070 Ti. Solo mining with the official 1.6.7 node binary accepted 1,016 pre-H16 blocks and recovered after two forced connection drops.
The H16 checks used controlled jobs with the official gate and consensus library; they are not a claim of live mainnet H16 acceptance before activation. Full validation details are in docs/H16.md.
Packages: Linux NVIDIA modern/legacy, Linux AMD, HiveOS, mmpOS, SMOS, Windows NVIDIA/AMD, macOS arm64, and Docker ocminersupr/keryx-miner-supr:0.16.0 / latest. Keep the executable and its bundled libraries together. Existing project and third-party licenses are preserved.
Metal H15/H16 remains unsupported. Windows GPU hardware, physical AMD H16 transitions and physical Pascal/Volta H16 transitions were not tested for this release. Prior CMP full inference/replay qualification remains incomplete after its thermal/power guard stop; its mining correctness and measured cache gain passed.