Skip to content

gkigiermo/prize-gpu-fpga-msm

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Accelerating MSM Operations on GPU/FPGA

Prize Sponsor: Aleo

Prize Architect: Aleo

Prize Description

Summary

Multiscalar multiplication (MSM) operations are an essential building block for zk computations. This prize will focus on computing the fastest MSM using either GPUs or FPGAs.

Optimization Objective

Compute the fixed-base point MSM for 2^26 randomly sampled scalars from the BLS 12-377 base field with lowest latency .

More specifically, given a fixed set of elliptic curve points from the BLS 12-377 G1 curve:

and a randomly sampled input vector of finite field elements from the scalar field:

Calculate elliptic curve point Q when n = 2^26

Constraints

  • Sufficient documentation must be provided along with the implementation

  • Only a single GPU/FPGA may be used; the problem may not be parallelized across multiple hardware instances.

  • Submissions may be written in any language. The provided test harness, however, will be in Rust. So competitors submitting solutions using other languages will be required to create their own Rust bindings

Timeline

June 1 - Final competitor selection

June 10 - Competition begins

July 25 - Mid-competition IPR

September 10 - Deadline for submissions

October 1 - Winners announced


Judging

Competitors for the prize will be selected based on their prior documented experience and academic achievement.

Submissions will be analyzed for both correctness and performance.

Correctness

We will provide a set of test input/output vectors so that the competitors can sanity check the correctness of their code.

The final correctness of the submission will be tested using randomly sampled test inputs/outputs that are not disclosed to the competitors during the competition. Submissions that fail any test cases will be judged as incorrect and lose the opportunity to win the prize.

Performance

Given input vectors consisting of 2^26 fixed elliptic curve points (bases), participants will compute a scalar multiplication of those bases with a set of four vectors of scalar elements from the associated BLS 12-377 G1 field in succession. Competitors will be provided with a set of test vectors to use while building the solution (of course, competitors can, and are encouraged to, use other vectors during the design and build process).

For scoring, solutions will be run using four randomly selected test vectors as input across ten trials in total. The winning submission will be the one with the lowest average latency across all ten trials.

In addition, all submissions will be manually reviewed by the prize committee appointed by the prize sponsor, Aleo.

Hardware & Benchmarks

Competitors will be given access to one of the following:

  1. A dedicated instance of baseline image consisting of an AMD Epyc Milan CPU (8 cores), an A40 NVIDIA GPU (w/ 80 GB PCIe connection), and 48 GB of RAM provided by Coreweave

  2. AWS credits to use an Amazon F1 instance FPGA f1.2xlarge consisting of a single FPGA and 8 CPU cores.

In addition, the prize sponsor (Aleo) will provide a test harness consisting of test vectors as well as a reference implementation that consists of:

  • A starting GPU library for MSM (on NVidia cards) provided by Aleo

  • A modified version of the ZCash FPGA MSM library for F1

Prize Allocation

All submissions that receive a prize must beat the benchmark solution.

Top GPU solution 25% of total prize
Top FPGA solution 25% of total prize
Most creative solution 10% of total prize

In addition the winning submission has the opportunity to earn up to 40% of the total prize amount based on how much their solution improved on the starting benchmark. See the table below:

2x better than baseline +10%
3x better than baseline +10%
5x better than baseline +20%

Example: a competitor submits a solution using a GPU that is the fastest overall solution and beats the provided benchmark implementation by 5x. That competitor would earn 20% (for being the top GPU submission) plus 10% (for beating the baseline by 2x) plus 10% (for beating the baseline by 3x) and another 10% for beating it by 5x. So the total share of the prize money would be 25% + 5% + 15% + 20% = 65%

Prizes will be given out in good faith and in the sole discretion of the prize committee.

Notes

All submission code must be open-sourced at the time of submission. Code and documentation must be dual-licensed under both the MIT and Apache-2.0 licenses.

Questions

If there are any questions about this prize, please contact Alex at Aleo: zprize@aleo.org

References

[1] Scalar-multiplication algorithms. https://cryptojedi.org/peter/data/eccss-20130911b.pdf

About

Accelerating MSM Operations on GPU/FPGA

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Rust 68.0%
  • Shell 32.0%