How Fast Can FlashProve Go?

The author estimates an upper bound on the throughput of the FlashProve ZK-SNARK for LLM inference, concluding that performance without protocol changes is capped at around 2,800 tokens per second per H100. This limitation is primarily driven by the costs of Reed-Solomon encoding and BLAKE2b Merkle hashing required for intermediate commitments. Even under this upper bound, performance remains about tenfold below the limits imposed purely by the model's floating-point operations. Open article →