/fpga_snark_prover

An acceleration engine for proving SNARKS over the bn128 curve, targeted for AWS FPGAs

Primary LanguageSystemVerilog

FPGA SNARK prover targeting the bn128 curve.

This repo contains the results of an Ethereum project into developing a FPGA SNARK prover (still in progress).

Project goals

The goal of this project is to accelerate zk-SNARK proof creation by offloading operations to a FPGA:

  • FPGA acceleration of bn128 curve arithmetic
    • Implement base curve arithmetic in Montgomery form, kernel architecture
    • Implement G1 (Fp) & G2 (Fp2) point multiplication
    • Implement architecture on AWS and show functionality in the Vitis accelration framework
  • Multi exponentiation (multiexp) - implement G1_MULTIEXP and G2_MULTIEXP
    • Modify point multiplication architecture from #1 on FPGA utilizing some method to accelerate multiple parallel point multiplications (e.g. utilize window / NAF / pre calculation)
    • Implement G1_MULTIEXP
    • Implement G2_MULTIEXP
  • QAP evaluation - (i)FFT - implements CALC_H and all sub operations
    • Implement polynomial field arithmetic on FPGA
    • Implement the FFT / iFFT on FPGA

Getting started

Running at home on a FPGA

  1. Clone this repo locally and include the RTL code into the top level for your board (only tested on Xilinx Ultrascale+ parts). Top level files are in fpga_snark_prover/src/rtl/.
  2. Run this on the Amazon AWS infrastructure. Instructions are in the example projects are here.

FPGA top level

Folder structure

fpga_snark_prover/  -- Files directly related to this project.
ip_cores/           -- RTL ip cores for general use in this project.
submodules/         -- Git submodules that reference other repositories.

Each goal will have a corresponding RTL top level and source in the fpga_snark_prover folder which can be ran separately, as well as a RTL kernel pre-compiled (along with instructions to compile the binary if needed). The kernels will be able to run in a Amazon AWS Vivado Vitis (SDAccel) environment as accelerators, interfaced to user logic (look in fpga_snark_prover\kernel to see this). The kernels have access to 4x 16GB DDR4 banks when run on Amazon AWS F1 FPGA instances.

More information on FPGA kernels can be found here in the Xilinx Vitis user guide here.

You can also use Vivado to modify and simulate the RTL by creating a project and adding all the sources from each folder, and then running the testbenches in the /tb/ folders.