Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EPIM: Efficient Processing-In-Memory Accelerators based on Epitome

Chenyu Wang, Zhen Dong, Daquan Zhou, Zhenhua Zhu, Yu Wang, Jiashi Feng, Kurt Keutzer

TL;DR

This work addresses the challenge of deploying large CNNs on Processing-In-Memory (PIM) accelerators by introducing EPIM, an epitome-based neural operator tailored for PIM hardware. EPIM replaces standard convolutions with compact epitomes, incorporates a PIM-aware layer-wise design, and uses epitome-aware quantization plus hardware data-path adaptations (IFAT, IFRT, OFAT) to reduce memory and compute demands. Layer-wise epitome design driven by evolutionary search and output channel wrapping mitigate latency and energy overheads, achieving up to 30.65× crossbar compression with modest accuracy loss (e.g., 71.59% top-1 on ImageNet for 3-bit EPIM-ResNet50) and substantial energy savings. The results demonstrate that epitome-based operators can outperform pruning on PIM by offering strong compression-accuracy-energy tradeoffs, enabling practical deployment of large CNNs on PIM platforms.

Abstract

The utilization of large-scale neural networks on Processing-In-Memory (PIM) accelerators encounters challenges due to constrained on-chip memory capacity. To tackle this issue, current works explore model compression algorithms to reduce the size of Convolutional Neural Networks (CNNs). Most of these algorithms either aim to represent neural operators with reduced-size parameters (e.g., quantization) or search for the best combinations of neural operators (e.g., neural architecture search). Designing neural operators to align with PIM accelerators' specifications is an area that warrants further study. In this paper, we introduce the Epitome, a lightweight neural operator offering convolution-like functionality, to craft memory-efficient CNN operators for PIM accelerators (EPIM). On the software side, we evaluate epitomes' latency and energy on PIM accelerators and introduce a PIM-aware layer-wise design method to enhance their hardware efficiency. We apply epitome-aware quantization to further reduce the size of epitomes. On the hardware side, we modify the datapath of current PIM accelerators to accommodate epitomes and implement a feature map reuse technique to reduce computation cost. Experimental results reveal that our 3-bit quantized EPIM-ResNet50 attains 71.59% top-1 accuracy on ImageNet, reducing crossbar areas by 30.65 times. EPIM surpasses the state-of-the-art pruning methods on PIM.

EPIM: Efficient Processing-In-Memory Accelerators based on Epitome

TL;DR

This work addresses the challenge of deploying large CNNs on Processing-In-Memory (PIM) accelerators by introducing EPIM, an epitome-based neural operator tailored for PIM hardware. EPIM replaces standard convolutions with compact epitomes, incorporates a PIM-aware layer-wise design, and uses epitome-aware quantization plus hardware data-path adaptations (IFAT, IFRT, OFAT) to reduce memory and compute demands. Layer-wise epitome design driven by evolutionary search and output channel wrapping mitigate latency and energy overheads, achieving up to 30.65× crossbar compression with modest accuracy loss (e.g., 71.59% top-1 on ImageNet for 3-bit EPIM-ResNet50) and substantial energy savings. The results demonstrate that epitome-based operators can outperform pruning on PIM by offering strong compression-accuracy-energy tradeoffs, enabling practical deployment of large CNNs on PIM platforms.

Abstract

The utilization of large-scale neural networks on Processing-In-Memory (PIM) accelerators encounters challenges due to constrained on-chip memory capacity. To tackle this issue, current works explore model compression algorithms to reduce the size of Convolutional Neural Networks (CNNs). Most of these algorithms either aim to represent neural operators with reduced-size parameters (e.g., quantization) or search for the best combinations of neural operators (e.g., neural architecture search). Designing neural operators to align with PIM accelerators' specifications is an area that warrants further study. In this paper, we introduce the Epitome, a lightweight neural operator offering convolution-like functionality, to craft memory-efficient CNN operators for PIM accelerators (EPIM). On the software side, we evaluate epitomes' latency and energy on PIM accelerators and introduce a PIM-aware layer-wise design method to enhance their hardware efficiency. We apply epitome-aware quantization to further reduce the size of epitomes. On the hardware side, we modify the datapath of current PIM accelerators to accommodate epitomes and implement a feature map reuse technique to reduce computation cost. Experimental results reveal that our 3-bit quantized EPIM-ResNet50 attains 71.59% top-1 accuracy on ImageNet, reducing crossbar areas by 30.65 times. EPIM surpasses the state-of-the-art pruning methods on PIM.
Paper Structure (21 sections, 7 equations, 4 figures, 3 tables, 1 algorithm)

This paper contains 21 sections, 7 equations, 4 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. Process-In-Memory Architecture for CNN Acceleration
  4. Epitome
  5. Quantization
  6. Framework Overview
  7. EPIM Design Method
  8. Mapping and Adjusting Epitomes for PIM Accelerators
  9. Quantization Method for Epitome
  10. Architecture and Data Path of EPIM
  11. EPIM Design Optimization
  12. Motivation
  13. Layer-Wise Epitome Design
  14. Output Channel Wrapping
  15. Experiments
  16. ...and 6 more sections

Figures (4)

  • Figure 1: Epitome's Reconstruction into Convolution
  • Figure 2: Illustration of EPIM Framework.
  • Figure 3: The parameter size, latency, and energy consumption correspond to different layers (Layer 9, 41, 67) in ResNet50, with or without epitome.
  • Figure 4: Comparison of the uniform-epitome solution with two optimization methods: EPIM-Channel Wrapping and EPIM-Evo-Search, each used individually. EPIM-Opt refers to a combination of these two methods.