Fermah is shared ZK proving infrastructure, plus a Matchmaker that routes proofs to the least-bad GPU.
4 comments
So its basically proof outsourcing with a scheduler? Feels like the hard part is GPU heterogeneity and queueing, not the ZK side.
GPU heterogeneity and queueing are the hard part, ZK mostly just sets the constraints.
If the Matchmaker sends a proof to the "least-bad GPU", what stops a slow or faulty prover from just sitting in the queue and making the user wait, do you need some kind of prepayment or slashing for that?
The queueing concern is real, but slashing only helps if you can define a clean fault (missed deadline, bad output), it does not make a slow prover fast, and a timeout plus reassign is probably the more useful lever here....
Compared with Risc Zero Bonsai, Fermah sounds more flexible on capacity sourcing (especially if it can pick around GPU heterogeneity), but Bonsai's tighter control over the proving stack seems easier to police when you care about who actually stalled.