TORONTO — Resistive random access memory (ReRAM) and other emerging memory technologies have been getting a lot of attention in the past year as semiconductor companies look for ways to more efficiently deal with the requirements of artificial intelligence and neuromorphic computing.
At the International Electron Devices Meeting (IEDM) in San Francisco earlier this month, there were several papers presented that dealt with using emerging memory in neomorphic computing from companies the likes of IBM and various universities.
"The neuromorph crowd is excited about this type of thing," said Jim Handy, a veteran memory market watcher who is principal analyst at Objective Analysis. "I wouldn't say that any [the emerging memory technologies} stands out. They all have something. The question is who is going to get something meaningful to the market first."
Neuromorphic applications are designed to specifically mimic how the human brain learns and processes information, and ReRAM devices show promise for enabling high-density and ultimately scaled neuromorphic architectures because they are significantly smaller and more energy-efficient than current AI data centers. They also mimic the brain’s biological computation at the neuron and synaptic level.
"The real beauty is that it could dramatically reduce the price of an AI system and the power consumption," Handy said.
To that end, Weebit Nano recently partnered with the Non-Volatile Memory Group of the Indian Institute of Technology Delhi (IITD) on a collaborative research project that will apply Weebit’s SiOx ReRAM technology to computer chips used for AI.
IITD’s Non-Volatile Memory Research Group is led by Professor Manan Suri under the university’s Department of Electrical Engineering. He said that the group is driven by a vision of creating an “NVM-centric future” in which the role of NVM goes beyond simple storage. Current research areas include computing, sensing, and security. The university also runs specialized courses on volatile and non-volatile memory, and there are researchers working on NVM-related materials in the university’s physics department. It’s also collaborating with IBM on AI initiatives.
ReRAM and other emerging NVM devices offer several benefits such as non-volatility, CMOS compatibility, ultra-high density, simple integration and fabrication, and low cost, said Suri, and NVM devices are important for realizing AI hardware, accelerators, and neuromorphic hardware applications. “Memory can pretty much make or break a dedicated neuromorphic AI hardware system,” he said. “It’s important to choose state of the art.”
IITD often partners with industries and startups for applied and exploratory research. Suri said that Weebit is among the few companies that is consistently progressing with its ReRAM and is open to collaboration with universities. In the short term, IITD expects the research to lead to effective realization of high-density non-volatile neuromorphic circuits. Longer term, there will be a need for more and more dedicated AI/neuromorphic chips due to the ever-increasing amount of data, he said. “Dedicated neuromorphic/AI hardware holds the promise of tackling the enormous data problem in a more energy-efficient and sustainable way compared to general-purpose CPUs.”
Weebit’s research efforts don’t change its commercialization plans for its SiOx ReRAM technology, but it wants to make sure that it leverages its full capability across advanced applications, said CEO Coby Hanoch. Collaborating with IITD is a cost-effective way of keeping its technology on the leading edge, and he sees other similar partnerships as an important element of the company’s future going forward. Having technology that can be sampled was an important first step. “This is the first real project that we’re engaging with but definitely not the last one," said Hanoch. "We’re already talking to others.”
Weebit CTO Amir Regev said that with AI starting to permeate the everyday lives of the average person, it’s essential to have technologies that can help it scale more quickly. ReRAM, in particular, makes sense for AI applications because it can stay small, and from a power consumption perspective, it works well because it accommodates the human-brain–like spikes that occur in neuromorphic applications. And because it’s early days for ReRAM in AI, he said, there are no set standards yet. “It’s kind of a green field. It’s very innovative, and many researchers are doing other, different stuff.”