mimblewimble team mailing list archive

[John Tromp <john.tromp@xxxxxxxxx>]

dear John,

> Thanks for the update. Just curious, what's the time tradeoff for that 20x
> memory? This optimization becomes difficult, especially on GPU, over Cuckoo
> 30, correct?

The speedup is about 4x on a cpu. But since the solver becomes more similar
to an Equihash solver (both dominated by bucket sorting), we can expect a
similarly large gap between CPU and GPU of 10x-20x.

Assuming we're happy with a mining landscape dominated by GPU cards
with at least 4GB, then cuckoo30 is a preferred size, since it should take
about 3GB of memory. Devices with much less memory available will have
the option to use the current high latency miner, which needs less than 200MB,
but for these mining will not be progress-free (they will manage only
one or very few attempts per block interval, the last one of which
will be wasted).

regards,
-John