Halo2 batch MSM, about 2x faster
Strahn Cryptography said it made Halo2 batch same-base MSM about 1.9-2x faster than pasta_curves on an Apple M4.
The Strahn team shared on the forum that it achieved a verifiable 1.9-2x speedup in Halo2 batch MSM (multi-scalar multiplication) over the industry-standard pasta_curves crate. Measured on a standard M4, at k=11 pasta_curves took 34.6ms and its own implementation (fuji batch) 17.0ms, 2.0x; k=12 was 66.9ms versus 33.4ms, 2.0x; k=13 was 110.5ms versus 59.3ms, 1.9x; and k=14 was 223.2ms versus 117.8ms, 1.9x. At k=8 there was no benefit because of cache warm-up overhead.
Technically, it used NEON/SIMD registers like a scratchpad to parallelize operations across the Apple CPU's four execution ports, and shared computation over the same G-base to raise overall speed. The code can be found in the StrahnIo/GoldenDelicious and StrahnIo/FUJI repositories on GitHub.
Sean Bowe replied on X that they would like to look at this work, saying Tachyon also uses the Pasta curves and that most of its proving cost is still MSM. Sean explained that halo2 proofs for Orchard and Ironwood spend a lot of time on FFT, while Tachyon's Ragu relies only on MSM without FFT. The author took this to mean the speedup could be larger in Tachyon. The author is currently waiting for Sean's review, and ends the post by asking the community whether they should apply for a grant to extend this to iPhone A-series chips.
This summary is based on the original; check the original for exact details.