OISMA: In-Memory Stochastic Multiplication for AI Matrix Workloads
A new in-memory computing architecture called OISMA (On-the-fly In-memory Stochastic Multiplication Architecture) has been proposed to address the computational bottleneck of matrix-multiplication workloads in artificial intelligence models. OISMA leverages the bent-pyramid (BP) quasi-stochastic computing system to perform in situ stochastic multiplication during normal memory read operations, with negligible additional cost. The architecture aims to overcome limitations of both digital/binary-based and analog IMC architectures, which suffer from performance and energy efficiency degradation. An accumulation periphery accumulates the output multiplication bitstreams to achieve matrix operations. The work is detailed in arXiv:2508.08822v2.
Key facts
- OISMA is an energy-efficient IMC architecture.
- It uses the bent-pyramid (BP) quasi-stochastic computing system.
- It converts normal memory read operations into in situ stochastic multiplication.
- The accumulation periphery accumulates output multiplication bitstreams.
- It targets matrix-multiplication workloads in AI models.
- It aims to avoid the von Neumann bottleneck.
- It addresses limitations of digital/binary and analog IMC architectures.
- The paper is on arXiv with ID 2508.08822v2.
Entities
Institutions
- arXiv