Facebook Yann Lecun: RISC-V will explode in edge AI

Facebook's chief AI scientist Yann Lecun said a boom in low-cost edge AI chips using RISC-V technology is coming.

Yann Lecun said that with Nvidia's acquisition of Arm, more neural network accelerators for edge applications will consider using RISC-V, Facebook's chief artificial intelligence scientist said at the Innovation Day hosted by CEA-Leti.

He said: "The industry is changing. The Nvidia and ARM deal made people uneasy, but the emergence of RISC-V allows the industry to have independent cores and NPUs (neural processing units). These products are very cheap, less than $10, and many are outside of China, and these products will become ubiquitous. Will RISC-V take over the world through this field?"

He discredited a major Leti project working on spiking neural networks and RRAM, among other things, that takes an analog circuit approach.

“The main problem with analog implementations is that it’s hard to multiplex hardware with analog neural networks,” he said. “When you do convolutions and reuse hardware, you have to multiplex the hardware, so you have to have a way to store the results, and then you need analog memory or ADC and DAC converters, which kills the whole matrix. So, unless we have cheap low-power analog memory, it’s not going to work. I doubt it, maybe memristor arrays or spintronics, but I kind of doubt it.”

"There is no doubt that edge AI is a very important topic," he said. "In the next two or three years, it will not be an exotic technology, but to reduce power consumption as much as possible, prune neural networks, optimize weights, and shut down unused parts of the system." Yann Lecun said, "Our goal is to use AR devices with chips in the next 2 to 3 years, and this goal is coming."

“Ten years from now, what breakthroughs will there be without analog hardware multiplexing?” he asked. “Whether we can come up with a scheme to reduce the size so much for a single chip without data shuffling and hardware multiplexing is a major challenge.”

"Companies are developing 1nm and 2nm technologies for the next generation of chips, and I firmly believe that we can do different things for future hardware through sensors, neural networks and controllers," said Emmanual Sabonnadiere, CEO of Leti. "We are working hard to develop national plans and add scientific components to political decisions. Edge AI is about stopping data flooding and data privacy so that people can own their own data."

Leti is also part of the European Neural Network Initiative, which looks at new platforms for neural network chips.

“There is a new generation of technology to be worked on,” said Jean-Rene Lequeypes, deputy CEO and CTO of CEA-Leti. “Now we have more than 2,000 people working on the next generation of platforms. Imec in Belgium, Fraunhofer in Germany and Leti are developing a platform for edge AI, and in addition to that we are developing products at Inria in Grenoble, which I think Facebook and the big Silicon Valley companies need.”

The challenge is to integrate all the different elements without having to use extreme ultraviolet lithography at 5nm and below.

“We want to get to the ultimate performance of 1,000TOPS/mW, which is a very big challenge, as well as how to use memory, different technologies and how to integrate them without having to use EUV,” Lequeypes said.