Content
@
0 reply
0 recast
0 reaction
Sam Ragsdale
@samrags
SNARKs on ARM just got ~10% faster. The primary operation in a SNARK is mul between field elements – maybe 80% of total CPU time. These 50-70 assembly instructions run repeatedly on different inputs to create proofs. We upstreamed a PR to halo2curves to shorten this hot path.
2 replies
6 recasts
29 reactions
Sam Ragsdale
@samrags
Many SNARK libraries depend on halo2curves by the PSE team for their low level field and group arithmetic. halo2curves has excellent x86 acceleration but ARM – an increasingly relevant compile target – isn't yet optimized.
1 reply
0 recast
2 reactions
Sam Ragsdale
@samrags
The PR implements the fast CIOS method for Montgomery multiplication saving 7-15% on field multiplication time on ARM. https://github.com/privacy-scaling-explorations/halo2curves/pull/134 Thanks to the PSE team for the great library and rapid assist with the PR!
0 reply
1 recast
2 reactions
