Stanford AI 课程地图 EN GitHub ↗
← 返回课程地图
Stanford · 2026—27 CS 217 ML 系统与硬件 ★★★★☆

核心课程 · ML 系统与硬件

机器学习硬件加速器

Hardware Accelerators for Machine Learning (EE 244)

ML 硬件加速器:数据流、张量处理、存储层次、编译,以及 CNN/Transformer 的映射。

3–4 学分 2027 春 Letter or Credit/No Credit ★★★★☆
硬件加速器体系结构芯片

课程定位

CS 217(EE 244)由 Olukotun 和 Tambe 讲授,是 AI 芯片方向的经典课。课程从 CNN/Transformer 的计算模式出发,讲加速器架构、数据流、存储层次、低精度/稀疏和编译,最终做一个设计项目。

主要内容

  1. AI 计算模式:CNN、Transformer、训练 vs 推理
  2. 加速器架构:脉动阵列、张量核心、数据流
  3. 存储层次:片上缓存、HBM、数据复用
  4. 低精度与稀疏:量化、剪枝、压缩
  5. 编译与映射:算子调度、tiling
  6. 能效与面积:系统级 trade-off
  7. 设计项目

课程大纲 / 课件

#主题 / 讲者
01Introduction
02High-Level Synthesis
03Data Parallelism and Efficiency Metrics
04Quantization
05Systolic Arrays
06Vector and Scalar Units
07Memory Systems
08Interconnects and Network-on-Chip
09Compilers and Software Stack (Part 1)
10Compilers and Software Stack (Part 2)
11Review of Project Proposals (I)
12Review of Project Proposals (II)
13GPU Architectures
14Reconfigurable Dataflow Architectures
15SambaNova RDA (Raghu Prabhakar)
16Memory-Centric Computing (Part 1)
17Memory-Centric Computing (Part 2)
18AWS Inferentia and Trainium (Nafea Bshara)
19NVIDIA GPU Architecture for AI (Ronny Krashinsky)
20Poster Sessions

以 Winter 2026 课程表整理;slides 与 lab 材料在 Canvas,公开页只给主题和 reading。

先修

数字电路/计算机体系结构;VLSI 或 FPGA 经验有帮助。

考核与工作量

论文阅读 + final design project。具体以当季官网为准。

适合谁

想做 AI 芯片、加速器、体系结构的人;硬件背景想理解 ML 负载的人。

学习建议

先理解 roofline 和矩阵乘的分块。配合 CS 149/CS 349E。

相关课程

官方课程描述(英文原文)

This course explores the design, programming, and performance of modern AI accelerators. It covers architectural techniques, dataflow, tensor processing, memory hierarchies, compilation for accelerators, and emerging trends in AI computing. This course will cover modern AI/ML algorithms such as convolutional neural nets, and Transformer-based models / LLMs. We will consider both training and inference for these models and discuss the impact of parameters such as batch size, precision, sparsity and compression on the accuracy of these models. Students will become familiar with hardware implementation techniques for using parallelism, locality, and low precision to implement the core computational kernels used in ML. Students will develop intuitions to make system-level trade-offs to design energy-efficient accelerators. Students will read recent research papers and complete a final design project.