LOG-CIM: A 116.4 TOPS/W Digital Computing-In-Memory Processor Supporting a Wide Range of Logarithmic Quantization with Zero-Aware 6T Dual-WL Cell
Authors:
- 5 November 2023
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-3
Abstract
Recently, low-bit precision quantization for DNN has been proposed to reduce both data movement and computation power [1], [2]. This approach achieves high accuracy by employing logarithmic quantization (LOGQ) for input activation or weight rather than integer quantization (INT), enabling coverage of a wide range. Within Computing-In-Memory (CIM), LOGQ ensures that only 1 bit of the bit-serial CIM input is set to ‘1’, resulting in lower power consumption. However, a physical shift alignment within the CIM is required each time to compute all CIM inputs ‘1’ to allow the accumulation of bit-wise multiplications at once. Otherwise, as in previous CIM works [3]–[5], converting exponent CIM inputs into integer data leads to extended alignment cycle latency. In [5], it reduced the latency through the double-bit cells, but this approach wastes power as the bit-line (BL) pre-charge cannot be reused.Keywords
Funding Information
- IC Design Education Center
This publication has 4 references indexed in Scilit:
- A 28nm 64-kb 31.6-TFLOPS/W Digital-Domain Floating-Point-Computing-Unit and Double-Bit 6T-SRAM Computing-in-Memory Macro for Floating-Point CNNs2023
- A 28nm 16.9-300TOPS/W Computing-in-Memory Processor Supporting Floating-Point NN Inference/Training with Intensive-CIM Sparse-Digital Architecture2023
- ReDCIM: Reconfigurable Digital Computing- In -Memory Processor With Unified FP/INT Pipeline for Cloud AI AccelerationIEEE Journal of Solid-State Circuits, 2022
- DeepShift: Towards Multiplication-Less Neural Networks2021