Revolutionary technology seeks to break the “memory bottleneck,” greatly improving computer processing speeds
A team of UCLA electrical and computer engineering professors has received a research grant, worth up to $5.9 million, from the Department of Defense to develop revolutionary energy efficient computing systems that can process massive amounts of data at speeds much faster than currently possible.
The project is called Spintronic Stochastic Dataflow Computing, and the research team aims to demonstrate a reduction in energy use of 60 times less than current technologies in data-intensive computer tasks such as machine learning.
The project looks to address the “memory bottleneck,” a slowing down of computing speeds as information is shuttled between memory and processing chips. As computers are required to handle larger and larger amounts of data, the energy required to process that information has become a limiting factor in computing speed. The research will be led by three electrical and computer engineering professors: Sudhakar Pamarti, Puneet Gupta, and Kang Wang, who holds UCLA’s Raytheon Chair of Electrical and Computer Engineering. The team members bring expertise in nanotechnology, computing systems and design automation, and integrated circuit design.
“It used to be that processing speeds were the main hurdles to computing performance, but as those have improved, that progress has been outpaced by the arrival of big data,” Pamarti said. “The paradigm has been flipped and the processors can’t handle all this data without using a lot of energy for moving it around. This memory bottleneck is now arguably the biggest challenge for the design of new computing systems. We’re developing an innovative two-part approach to work around that issue.”
First, the researchers will use a type of computer memory technology that uses dramatically less energy than current technologies. Called “MeRAM,” it offers the best reported combination of energy, speed, and density among existing and emerging non-volatile memory technologies.
The second idea is “stochastic computing”. This requires very compact hardware that work in parallel with each other, which relaxes the memory bottleneck problem. Furthermore, stochastic computing makes computations progressively more accurate. This way, devices can be tuned to trade off their energy consumption for approximate answers when those will suffice. In addition, the two technologies will work very well together.
The grant is from the Defense Advanced Research Projects Agency, or DARPA. It is part of the organization’s larger effort, called the Electronics Resurgence Initiative, to solve fundamental challenges in computing and microelectronics to help continue Moore’s Law – which has successfully predicted the continual shrinking of transistors for microelectronics over the past five decades.