EigenArch: A Low Rank Hardware Machine Learning Accelerator
Abstract
The goal of the EigenArch project is to develop a faster and more energy efficient machine learning accelerator. A machine learning accelerator is a hardware system which can perform important artificial intelligence tasks more rapidly than a regular computer processor. The EigenArch accelerator will be based on analog technology (like resistive switches, phase change memories, etc.) rather than digital logic, because it offers the advantages of a design that unifies memory and computation to achieve orders of magnitude lower energy assumption and faster operation. However, due to their physics of operation and technological newness, these analog devices have imperfections and non-ideal behavior that prevent the system from learning efficiently. The proposed EigenArch will uniquely use decomposition techniques implemented in traditional CMOS technology to stabilize the training of novel neural networks and reduce the hardware overhead in neuromorphic systems. With EigenArch,emerging analog memory devices can be used despite their imperfections, bringing this new technology closer to deployment.The team, comprising of George Washington University, NIST and Western Digital Research have experience in the field of resistive switching devices, hardware modelling and prototyping and have invested significant resources already, including personnel, preliminary research, and developing test CMOS circuitry to prototype training with different kinds of switching devices. The mid-term deliverable is a working Verilog model of the EigenArch system used to determine potential challenges and explore key performance metrics. The end-term deliverable is a hardware emulator of this system, like an FPGA controller, with a 2T-1R analog memory / CMOS platform with 20k ReRAM devices.Based on the success of the project, commercialization can be achieved by moving the project from Western Digital Research to other divisions of Western Digital (WD) since there is a significant demand for machine learning products. WD has both domestic and overseas fabrication capabilities and partnerships to get this technology built and in use for DoD applications or to the market. With artificial intelligence having a transformative impact across many industries, including transportation, medicine, and manufacturing, staying a leader inmachine learning technology is important to U.S. competitiveness and national security.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Dec 16, 2019
- Source ID
- N000142012031
Entities
People
- Gina Adam
Organizations
- George Washington University
- Office of Naval Research
- United States Navy