According to our benchmarks, almost all time is spent on operations performed by the substrate-bn crate.
https://github.com/unionlabs/union/blob/8a809a05417df4eedb718404319f256ecb20adb3/lib/cometbls-groth16-verifier/Cargo.toml#L22
RiscZero provides some accelerated crates:
https://dev.risczero.com/api/zkvm/acceleration#adding-accelerator-support-to-crates
But judging by substrate-bn's Cargo.toml, this won't be an easy patch:
https://github.com/paritytech/bn/blob/master/Cargo.toml
More work needs to be done on investigating how we can accelerate this.