Targets

We are pushing for:

• 1Mb every 5 seconds (200Kb/s)
• 1Mb blob, which is roughly 32000 proofs regardless the shape of the encoded matrix

Current state

In our current implementation, we can compute proofs between 2.5ms and 1.3ms per proof, depending on the polynomial degree. This is in a highly parallelized environment.

This matches other reference implementation as rust-kzg:

This still leave us far away from our target as with this we can push for a range of 32-38Kb/s depending on the column configuration.

Options

We have available the fk20 (Feist-Khovratovich) method (amortized calculation). There are some benchmarks available here.

It should improve times *on single thread, *but it is not sure if it can be parallelized as the other method.

Some benchmarks throws:

computing proofs with fk20. POLY_SIZE 4096, NUM_INPUT_POINTS 2, NUM_PROOFS 4096
time:   [3.7427 s 3.7489 s 3.7555 s]

This results in a 1.09ms per proof (partially parallelized).

Even using this, for our current target we would still have to use about ~28seconds to compute all the needed proofs**. And, as it is not clear if it can be parallelized this values may be the limit.**