.png){kind=icon}
.png){kind=icon}
Return
Sparsity-Oriented MRAM-Centric Computing for Efficient Neural Network Inference
.png){kind=icon}
PRE
AI
DOI：10.1109/TETC.2023.3326312.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Near-memory computing (NMC) and in- memory computing (IMC) paradigms show great importance in non-von Neumann architecture. Spin-transfer torque magnetic random access memory (STT-MRAM) is considered as a promising candidate to realize both NMC and IMC for resource-constrained applications. In this work, two MRAM-centric computing frameworks are proposed: triple-skipping NMC (TS-NMC) and analog-multi-bit-sparsity IMC (AMS-IMC). The TS-NMC exploits the sparsity of activations and weights to implement a write-read-calculation triple skipping computing scheme by utilizing a sparse flag generator. The AMS-IMC with reconfigured computing bit-cell and flag generator accommodate bit-level activation sparsity in the computing. STT-MRAM array and its peripheral circuits are implemented with an industrial 28-nm CMOS design-kit and an MTJ compact model. The triple-skipping scheme can reduce memory access energy consumption by 51.5x when processing zero vectors, compared to processing non-zero vectors. The energy efficiency of AMS-IMC is improved by 5.9x and 1.5x (with 75% input sparsity) as compared to the conventional NMC framework and existing analog IMC framework. Verification results show that TS-NMC and AMS-IMC achieved 98.6% and 97.5% inference accuracy in MNIST classification, with energy consumption of 14.2 nJ/pattern and 12.7 nJ/pattern, respectively.
Keywords:
In-memory computing
near-memory computing
neural network
STT-MRAM
sparsity vector and matrix
Journal
IF:
5.4
Papers:
1.1K
Citations:
3.4K