Mining systems

Mine on the hardware AI actually uses.

Spinel's miner is planned as a native tensor-core implementation, not a generic hash loop wearing an AI label. The first implementation target is FP8 on Hopper-class hardware, followed by Blackwell and MXFP4.

Hardware roadmap

One protocol. Multiple accelerators.

Every fast path must reproduce the same protocol result as the independent reference implementation.

CUDA / HopperFirst

H100 FP8

Initial native mining target. Establish correct tensor execution, transcript folding, and end-to-end proof production.

SM90FP8CUDA
CUDA / BlackwellNext

B200 / RTX 5090

Extend FP8, then qualify the first MXFP4 work class against the same protocol reference.

FP8MXFP4CUDA
ROCm / CDNA4Planned

MI350-class

Cross-vendor qualification is a consensus requirement, not a benchmark side quest.

MXFP4ROCmORACLE PARITY

Reference miner interface

The exact CLI does not exist yet. This is the target operator experience:

$ spinel-miner run \ --pool testnet.pool.spinelresearch.net:1200 \ --wallet spin1... \ --worker h100-01 GPU 0 NVIDIA H100 class FP8_E4M3_R0 state mining work 742.1 Twork/s share accepted proof locally verified

The miner must locally verify candidate proofs before submission and report protocol work separately from vendor FLOPS.

Reference pool architecture

GPU miner ↓ jobs / shares Reference Stratum ↓ candidate proof Spinel full node ↓ accepted block Network Share accounting ≠ network blocks. Displayed throughput ≠ accepted work.

A reference Stratum implementation is planned before public testnet so third-party pools can integrate against a known-good protocol instead of reverse-engineering private behavior.

Mining metrics

Spinel intends to report the numbers operators actually need:

  • protocol work per second, by work class;
  • accepted and rejected pool shares;
  • stale work and generation changes;
  • power, temperature, and sustained utilization;
  • verifier cost per share and per candidate block;
  • accepted work per joule, not just peak tensor throughput.

Miner release gates

Gate A

CPU oracle exact

Public vectors establish the protocol result independent of GPU implementation.

Required
Gate B

Single-GPU parity

Random and adversarial FP8 cases produce identical proof inputs and target decisions.

Required
Gate C

Accepted devnet blocks

Miner, pool, and independent nodes agree on real candidate blocks.

Required
Gate D

Open public testnet

Source and binaries are released before any economically meaningful launch.

Required