> Interesting comparison results, I was originally surprised they are so close because for Ed25519 we've found that the sweet point was at about 40x, but here it's only 4x. I can imagine the results makes sense because in ECVRF we use
map to point (which doesn't exist in ed25519)
operations over non group generators (every H(m) and ChaumPed proofs result to unknown points) (and for ed25519 we applied precomputed tables which improve performance by far)
Ristretto implementations are slightly more expensive than pure ed25519 in Rust.
we don't utilize aG + bP double scalar mul optimizations.
There are a multiple different operations and libraries in play which makes it hard to compare them, so I suggest that we create benchmarks for basic operations (hash-to-curve, arithmetic) for the groups we support to make comparison easier.
We don't utilise double scalar multiplication, but I'll check if it's possible.
There are a multiple different operations and libraries in play which makes it hard to compare them, so I suggest that we create benchmarks for basic operations (hash-to-curve, arithmetic) for the groups we support to make comparison easier.
We don't utilise double scalar multiplication, but I'll check if it's possible.
Originally posted by @jonas-lj in https://github.com/MystenLabs/fastcrypto/issues/381#issuecomment-1411609755