ZKP16 Hardware Acceleration of ZKP

ZKP学习笔记

ZK-Learning MOOC课程笔记

Lecture 16: Hardware Acceleration of ZKP (Guest Lecturer: Kelly Olson)

  • The What and Why of Hardware Acceleration

    • Hardware acceleration is the use of dedicated hardware to accelerate an operation so that it runs faster and/or more efficiently.
    • Hardware acceleration can involve optimizing functions and code to use existing hardware (COTS) or it may involve the development of new hardware designed for a specific task.
      • COTS (commercially available off-the-shelf) hardware includes CPUs, GPUs, and FPGAS
      • Custom hardware is often referred to as an ASIC
    • Examples
  • Hardware acceleration for crytpo

  • Why HW acceleration for ZKP

    • ZK (and non-ZK) proof generation has high overheads relative to native computation
  • Goals of HW acceleration for ZKP

    • Throughput: increase the number of operations per system
    • Cost: reduce the cost of operation e.g. Bitcoin mining rigs are designed to reduce capital expenses ($/hash) and operational expenses (watts/hash)
    • Latency: reduce the time of an individual operation e.g. 2kBridges may want to reduce the proof generation time for faster finality
  • Key Computational Primitives of ZKP

    • Each proof system, and associated implementation will have slightly different computational requirements.

    • Across a variety of proof systems these are three of the most computationally expensive operations

      • Multiscalar Multiplication (MSM)
        • A 'dot product' of elliptic curve points and scalars

        • Easily paralledizable

        • Optimization

          • When performing a MSM off of the host device, the scalars and sometimes points must be moved to the accelerator. The available communication bandwidth limits the maximum possible performance of the accelerator.
      • Number Theoretic Transformation (NTT)
        • Common algorithms like Cooley-Tukey reduce complexity from O ( N 2 ) O(N^2) O(N2) to O ( N I o g N ) O(NIogN) O(NIogN)
        • Not Easily paralledizable
        • Furthermore, these elements must be kept in memory to be operated on, imposing high memory requirements
      • Arithmetic Hashes (e.g., Poseidon)
    • SNARK V.S. STARK

      • The MSM, NTT and Hashes take 2/3 or more time in the proving system
    • Foundational Primitive: Finite Field Arithmetic (especially ModMul)

  • Hardware Resources Required

    • Determining Computational Cost

    • Selecting the Right Hardware

      • Given that these workload are driven predominately by modular multiplication, we should look for platforms can perform a large number of multiplications, quickly and cheaply
      • Estimated HW performance can be evaluated by looking at # of hardware multipliers, size of hardware multipliers, and speed/frequency of each instruction
      • Examples
    • Two Key Components to HW Acceleration

      • 'HW friendly' Algorithm
      • Efficient Implementation
  • Limits of Acceleration

    • Acceleration Pitfalls

    • Production Examples: Filecoin

  • Current Status of Hardware Acceleration

  • Future Directions for Hardware Acceleration

相关推荐
lingggggaaaa39 分钟前
小迪安全v2023学习笔记(八十二讲)—— Java组件安全&Solr&Shiro&Log4j&CVE复现
笔记·学习·安全
好望角雾眠43 分钟前
第四阶段C#通讯开发-1:通讯基础理论,串口,通讯模式,单位转换,代码示例
开发语言·笔记·c#·串口·通讯
贝塔实验室2 小时前
两种常用的抗单粒子翻转动态刷新方法
论文阅读·经验分享·笔记·科技·学习·程序人生·fpga开发
范纹杉想快点毕业9 小时前
ZYNQ PS 端 UART 接收数据数据帧(初学者友好版)嵌入式编程 C语言 c++ 软件开发
c语言·笔记·stm32·单片机·嵌入式硬件·mcu·51单片机
茯苓gao11 小时前
STM32G4 电流环闭环
笔记·stm32·单片机·嵌入式硬件·学习
easy202011 小时前
机器学习的本质:从跑模型到真正解决问题
笔记·学习·机器学习
汇能感知12 小时前
摄像头模组在智能家居设备中的应用
经验分享·笔记·科技
shizidushu13 小时前
Graph RAG论文阅读笔记
论文阅读·笔记·graphrag
0_0梅伊阁诗人15 小时前
Django ORM 模型
开发语言·数据库·笔记·python·oracle·django
m0_5782678615 小时前
从零开始的python学习(九)P142+P143+P144+P145+P146
笔记·python·学习