2/9 Full article:

3/9 Isolated benchmarks comparing autoprecompiles to existing precompiles in @openvm_org: - Keccak, U256: faster than manual precompiles! - SHA: autoprecompiles achieve close performance to manual. - EC/ECDSA: manual precompiles are faster, likely more hints could close the gap.

4/9 When benchmarking Ethereum block verification, we consider two scenarios: - Use autoprecompiles next to existing manual precompiles. - Do not use any manual precompiles.

5/9 In the first case, autoprecompiles already provide a 1.37x advantage over using vanilla OpenVM, including recursion!

6/9 We stop at 100 autoprecompiles above due to restrictions on recursion, which could be solved by a different proving system. We report results without recursion also for larger numbers of autoprecompiles. With 1000 autoprecompiles, we achieve a 1.84x speedup!

7/9 Finally, we removed all OpenVM precompiles from the guest program, simulating a world where they don't exist. We achieve a 2.62x improvement (with recursion) over software. As ECDSA is a major workload when verifying Ethereum blocks, OpenVM precompiles achieve better times.

8/9 Our experiments hint that manual precompiles can be made mostly obsolete sooner than later, eliminating a huge amount of engineering effort and security risk.

9/9 ZK proofs for Ethereum block verification: - 1.37x improvement today with recursion. - 1.84x improvement on app proofs. This will keep improving and we’re confident that we can eventually double the proving speed of Ethereum block verification with autoprecompiles alone!