Jump Crypto Security Engineer @0x0ece recently presented on FPGA Acceleration of Multi-Scalar Multiplication on a live stream hosted by @Scroll_ZKP. Watch the video below: youtu.be/aZ3CzKZBK38

Hear what @rahulmaganti_ had to say on Manta Network’s OpenZL Talk 10: CycloneNTT — Improving Twiddle Access for Number Theoretic Transforms. xitter.cf/MantaNetwork/status/16…

13/ The code will be open-sourced and available soon! If you want to chat ZK hardware acceleration, reach out to @rahulmaganti_, @0x0ece, and @nickraystalder. xitter.cf/IACR_News/status/15973…

12/ We explored architectures that process multiple layers at a time. Experiments with 1, 3, and 6-layers showed that the 3-layers approach yields the best latency vs. power tradeoff.

11/ Even without specialized hardware, this design also yields a 10-15% improvement over the current Plonky2 implementation. Check out the PR below. xitter.cf/_bfarmer/status/157518…

10/ At each layer, we compute several butterflies in parallel. Instead of pushing the output back into the same contiguous piece of memory, we modulate the outputs between a set of FIFOs to access the inputs for the next layer more easily.

9/ To manage data-flows and memory accesses to the twiddles more efficiently, we construct a set of FIFOs / queues.

8/ We developed a hardware-optimized quasi-streaming version of this algorithm for the FPGA.

7/ In our search for faster, memory-optimized FFT algorithms, we ran across a paper co-authored by none other than Jump’s Chief Science Officer, Kevin Bowers! ieeexplore.ieee.org/document…

6/ While our architecture is flexible, we focused on NTTs over the Goldilocks field. The Goldilocks field is a 64-bit field that enables efficient modular reductions on 64-bit machines. Computing a single reduction is equivalent to multiplying by a 32-bit integer.