Efficient SRAM-PIM Co-design by Joint Exploration of Value-Level and Bit-Level Sparsity

TL;DR

This paper introduces DB-PIM, a co-design framework that exploits value-level and bit-level sparsity in SRAM-PIM architectures, achieving substantial speedup and energy savings for neural network processing.

Contribution

The paper presents a novel algorithm-architecture co-design framework, DB-PIM, that effectively harnesses both value-level and bit-level sparsity in SRAM-PIM systems, improving efficiency.

Findings

Achieves up to 8.01x speedup in processing.

Provides 85.28% energy savings.

Demonstrates significant efficiency improvements in SRAM-PIM.

Abstract

Processing-in-memory (PIM) is a transformative architectural paradigm designed to overcome the Von Neumann bottleneck. Among PIM architectures, digital SRAM-PIM emerges as a promising solution, offering significant advantages by directly integrating digital logic within the SRAM array. However, rigid crossbar architecture and full array activation pose challenges in efficiently utilizing traditional value-level sparsity. Moreover, neural network models exhibit a high proportion of zero bits within non-zero values, which remain underutilized due to architectural constraints. To overcome these limitations, we present Dyadic Block PIM (DB-PIM), a groundbreaking algorithm-architecture co-design framework to harness both value-level and bit-level sparsity. At the algorithm level, our hybrid-grained pruning technique, combined with a novel sparsity pattern, enables effective sparsity management. Architecturally, DB-PIM incorporates a sparse network and customized digital SRAM-PIM macros, including input pre-processing unit (IPU), dyadic block multiply units (DBMUs), and Canonical Signed Digit (CSD)-based adder trees. It circumvents structured zero values in weights and bypasses unstructured zero bits within non-zero weights and block-wise all-zero bit columns in input features. As a result, the DB-PIM framework skips a majority of unnecessary computations, thereby driving significant gains in computational efficiency. Results demonstrate that our DB-PIM framework achieves up to 8.01x speedup and 85.28% energy savings, significantly boosting computational efficiency in digital SRAM-PIM systems.