Every row here is a CPU. Nockchain's GPU STARK provers are closed source and report rates in their own pool units, which cannot be checked against a challenge this server issued — so they are absent rather than unverified. Read this board as "what a CPU proves", not as the ceiling of the hardware.

The Logos AI-PoW puzzle (INT8 tiled matmul + recursive STARK certificate), verified the same way: every rate computed by this server from certificates it checked, of challenges it derived. Two statements share this board — dense, the single-tile benchmark, and canonical MoE, the block mainnet AI miners actually run — because MAC-equivalents is the unit consensus itself uses to compare work of different shapes. Tier is the expected grind attempts per win the run was measured at — clients pick it, so fast hardware can grind long enough to be measured rather than being drowned out by the fixed certificate-proving cost inside the window. Picking a tier cannot inflate a number: an easy tier earns less credit against that same fixed cost, and a hard one has to actually be ground out, attempt by verified attempt. CPU and GPU rows rank together: the device only searches for jackpots, and every win it finds is re-checked and certified on the CPU before this server verifies it.

These GPU figures are behind upstream. They were measured with the generic CUDA kernel at the nockchain commit this registry pins. Two days after that pin, upstream made a peak kernel the production default and reports the same card on the same statement going from 0.19 to 335 TMAC/s — about 1,700× — so treat the GPU rows here as a floor for the hardware, not a measurement of it. A re-pin and re-run is the next thing on the list. (Nockchain also admits mining inside real model inference, at throughput comparable to the optimised dedicated kernel; the interesting part there is that the compute is dual-use, not that it is faster.)