Close

Presentation

BRAHMS: Beyond Conventional RRAM-based Neural Network Accelerators Using Hybrid Analog Memory System
Time
Location
Event Type
Research Manuscript
Virtual Programs
Hosted in Virtual Platform
Keywords
AI/ML System Design
Topics
Design
DescriptionAccelerating Deep Neural Networks (DNN) with memristor-based processing-in-memory systems has been recognized as a promising approach. However, conventional accelerators are usually mixed-signal circuits with analog-digital converters (ADCs),which causes major performance degradation. In this work, we analyzed the invalid data flow between layers, and introduce the novel Analog Resistive Content Addressable Memories (ARCAM) into chip design to eliminate both non-linear function units and ADCs. BRAHMS utilizes the latest mapping schemes to achieve efficient transmission and conversion of analog signals in each layer.