Abstract
This letter presents an SRAM-based compute-in-memory (CIM) macro that uses 1-bit Δ Σ modulators to convert input and output activations to binary pulse waveform. The SRAM macro uses switched-capacitors for vector matrix multiplications and together with binary input activation improves linearity compared to current-domain SRAM CIM macros and allows reconfigurable activation resolution. The proposed macro is fabricated in 65 nm and benchmarked on MNIST and CIFAR-10 datasets with accuracies of 98.67% and 89.85%, respectively, with energy-efficiency in the range of 15.4-138.6 TOPS/W.
Original language | English (US) |
---|---|
Pages (from-to) | 293-296 |
Number of pages | 4 |
Journal | IEEE Solid-State Circuits Letters |
Volume | 6 |
DOIs | |
State | Published - 2023 |
Externally published | Yes |
Keywords
- Compute-in-memory (CIM)
- convolutional neural network (CNN)
- delta-sigma
- static random access memory
ASJC Scopus subject areas
- Electrical and Electronic Engineering