bf16: a 3.75× free speedup, and why it matters for miners
One change. Same model, same data, same seed, same config — except a single autocast wrapper in the training loop. The matmuls run in bf16 on hardware built for it.
The numbers
The loss barely moved: 3.8173 → 3.8163, a delta of 0.001 — well under the 0.0064 single-seed noise floor, i.e. statistically identical. So this is a 3.75× speedup with no accuracy cost. The H100 was doing roughly a quarter of its job under fp32.
Why it’s a protocol result, not just an ML one
Every submission in Ralph has to be re-trained inside the canonical proof test before it can be scored. At ~$13 per proof-test, miners can’t afford to iterate. At ~$3.50, they can. The noise floor tells a miner what ‘decisively beats the king’ means; this number tells them whether chasing that margin is affordable. Honest limit: this is still a single-box H100 number on FineWeb-Edu — not a claim about how the curve holds at 7B scale.
Honest limit
This is still a single-box H100 PCIe number on FineWeb-Edu sample-10BT. It says nothing about whether the same speedup curve holds at 7B scale — only that, at the scale Ralph evaluates submissions, the proof test got 3.75× cheaper with zero quality cost.