Sensitivity-Aware Mixed-Precision Quantization for ReRAM-based Computing-in-Memory
By Guan-Cheng Chen 1, Chieh-Lin Tsai 2, Pei-Hsuan Tsai 1, Yuan-Hao Chang 2
1 National Cheng Kung University, Tainan, Taiwan
2 National Taiwan University, Taipei, Taiwan
Abstract
Compute-In-Memory (CIM) systems, particularly those utilizing ReRAM and memristive technologies, offer a promising path toward energy-efficient neural network computation. However, conventional quantization and compression techniques often fail to fully optimize performance and efficiency in these architectures. In this work, we present a structured quantization method that combines sensitivity analysis with mixed-precision strategies to enhance weight storage and computational performance on ReRAM-based CIM systems. Our approach improves ReRAM Crossbar utilization, significantly reducing power consumption, latency, and computational load, while maintaining high accuracy. Experimental results show 86.33% accuracy at 70% compression, alongside a 40% reduction in power consumption, demonstrating the method's effectiveness for power-constrained applications.
To read the full article, click here
Related Semiconductor IP
- HBM4 PHY IP
- Ultra-Low-Power LPDDR3/LPDDR2/DDR3L Combo Subsystem
- MIPI D-PHY and FPD-Link (LVDS) Combinational Transmitter for TSMC 22nm ULP
- VIP for Compute Express Link (CXL)
- HBM4 Controller IP
Related Articles
- Pyramid Vector Quantization and Bit Level Sparsity in Weights for Efficient Neural Networks Inference
- From a Lossless (~1.5:1) Compression Algorithm for Llama2 7B Weights to Variable Precision, Variable Range, Compressed Numeric Data Types for CNNs and LLMs
- Synopsys 3DIO Solution for Multi-Die Integration (2.5D/3D)
- Functional Safety for Control and Status Registers
Latest Articles
- Making Strong Error-Correcting Codes Work Effectively for HBM in AI Inference
- Sensitivity-Aware Mixed-Precision Quantization for ReRAM-based Computing-in-Memory
- ElfCore: A 28nm Neural Processor Enabling Dynamic Structured Sparse Training and Online Self-Supervised Learning with Activity-Dependent Weight Update
- A 14ns-Latency 9Gb/s 0.44mm² 62pJ/b Short-Blocklength LDPC Decoder ASIC in 22FDX
- Pipeline Stage Resolved Timing Characterization of FPGA and ASIC Implementations of a RISC V Processor