Scalable Zero-Knowledge Proof Made Easy
summary
Zera extends OpenCilk/Kitsune/Tapir/LLVM to offload nested task-parallel C++ to GPUs, synthesize reductions, insert UVM prefetches and pool small allocations. Converting Shockwave zk-SNARK required 93 annotations on a 7,000-line codebase and yields 24× faster proof generation than a Rayon-parallel Rust baseline, coming within 1.6× of CUDA (manual CUDA added ~3,000 lines). Zera also parallelizes 3,000 lines of LAMMPS code and improves some large simulations up to 12%.
AI-written summary · about
chapters (12)
- 0:00Scalable ZK proofs intro
- 1:35What are ZK proofs
- 4:22Proof generation slowness
- 6:34Introducing Zera
- 13:04Understanding zk-SNARKs
- 18:02Five zk-SNARK patterns
- 20:29Zera extends OpenCilk
- 24:27CPU-GPU parallelism
- 26:43Efficient GPU memory
- 31:06CPU-GPU reductions
- 32:52Zera beyond ZKPs
- 34:43Summary
related talks
19mDIZK: A Distributed Zero Knowledge Proof System
24mZK: Accelerating DA Interop
13mUnveiling the Future of Zero-Knowledge Proofs
25mZK Rollup: Ethereum Scalability with ZKPs
38mNew Techniques for Zero-Knowledge: Leveraging Inefficient Provers to Reduce Assumptions, ...
48mzkSNARKs — Cryptographic Proofs for Scaling and Security Challenges
33mEducational Challenges in the Ever-Evolving Space of Zero-Knowledge
17mZK for the Enterprise - Quod Erit Demonstrandum (What Will Need to be Proven)