SK hynix and TetraMem have created a memristor-based in-memory computing SoC aimed at optimizing energy efficiency for AI edge devices. The architecture enhances performance for depthwise convolution while drastically reducing power consumption compared to traditional GPUs and NPUs, though it faces challenges in achieving competitive performance metrics.
SK hynix, TetraMem, and the University of Southern California have collaborated to develop a memristor-based system-on-chip (SoC). This chip is designed to improve the efficiency of AI inference tasks on edge devices while using significantly less power than GPUs and NPUs.
The architecture employs in-memory computing (IMC) techniques to carry out analog computations within memory arrays, reducing the need for data movement and thus conserving energy.
Specifically, the SoC integrates a conventional IMC crossbar architecture and a custom memristor-based design optimized for depthwise convolution (DWC), enhancing the performance of lightweight neural networks.
The SoC is powered by an embedded RISC-V processor and incorporates 10 neural processing units (NPUs), with one dedicated to optimizing depthwise convolution. Each NPU has a 256 x 256 memristor crossbar for analog vector-matrix multiplication, along with DACs and ADCs for data conversion.
The DWC-optimized NPU innovatively replaces traditional crossbar layouts with a zig-zag topology to boost efficiency and data processing.
In theoretical scenarios, the SoC could achieve a maximum performance of 2.54 TOPS, which remains significantly lower than requirements set by advanced systems such as Microsoft's Copilot+. Thus, while the chip demonstrates advancements in energy efficiency, its performance benchmarks may limit immediate practical applications in high-demand scenarios.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
SK hynix and TetraMem have created a memristor-based in-memory computing SoC aimed at optimizing energy efficiency for AI edge devices. The architecture enhances performance for depthwise convolution while drastically reducing power consumption compared to traditional GPUs and NPUs, though it faces challenges in achieving competitive performance metrics.