In an important speech at the Innovation Days at French research lab CEA-Leti, Yann LeCun, chief AI scientist at Facebook, mentioned that Nvidia’s acquisition of ARM could accelerate RISC-V to run neural networks for edge AI applications.

“The industry has changed, and the adoption of ARM, which belongs to Nvidia, makes people uneasy, but the advent of RISC-V makes people see a class with RISC-V cores and NPU (Neural Processing Unit) chips,” he said.

“These are incredibly cheap, less than $10, and a lot of them are outside of China, and they’re going to be everywhere.” “I wonder if RISC-V will take over the world there.”

He disapproved of one of Leti’s major plans to stimulate neural networks and similar methods such as Resistive RAM (RRAM), but the inventor of Convolutional Neural Networks (CNN) and Turing Award winner for AI There are other views.

“The main problem with analog implementations is that it is difficult to use hardware multiplexing with analog neural networks,” he said.

“When you do convolution and reuse hardware, you have to do hardware multiplexing, so there has to be a way to store the result, and then you need analog memory or ADC and DAC converters, which kills the whole idea. So unless we Have cheap low-power analog memory or it won’t work,” he said. “I’m skeptical, maybe a memristor array or a spintronic device, but I’m a little skeptical.”

“Certainly AI at the edge is a very important topic,” he said. “For the next two to three years, it’s not going to be a singular technology, it’s going to be about reducing power consumption as much as possible, pruning neural networks, optimizing weights. , shutting down unused parts of the system,” LeCun said. “Our goal is to bring that functionality to chips in AR devices within the next two to three years, and within five years to use such devices, and that’s coming,” he said. He says.

“Ten years from now, will there be some breakthroughs in spintronics, or any breakthroughs that allow analog computing without hardware multiplexing?” he asked. “Can we come up with such an idea? Without data shuffling and without hardware multiplexing, it’s a big challenge for a device like this to shrink dramatically in size for a single chip,” he said.

“Companies are developing 1nm and 2nm technologies for the next generation of chips, and I strongly believe that we can achieve the future of hardware with sensors, neural networks and controllers that will enable different developments,” said Emmanual Sabonnadiere, Leti’s CEO. “We are working hard to develop national plans and use science in political decision-making. Edge AI is designed to stop data flooding and data privacy so that people can own their data,” he continued.

Leti is also part of the European Neural Networks Initiative, which is working on a new platform for neural network chips.

“There is a new generation of technology being researched,” said Jean Rene Lequeypes, Deputy CEO and CTO of CEA-Leti.

“Right now, we have over 2,000 people working on next-generation technology. The challenge, he noted, is to integrate all the different components without having to use the extreme UV lithography required at 5nm and below. But we want to end up with 1,000TOPS/mW performance. , it’s a big challenge, but also how to use memory, different technologies and how to integrate them together without using EUV.”