RISC-V boom from edge AI says Facebook’s chief AI scientist
The move to RISC-V for running neural networks for edge AI applications is accelerated by the proposed takeover of ARM by Nvidia, says Yann LeCun, chief AI scientist at Facebook speaking at the Innovation Day of French research lab CEA-Leti.
“There is a change in the industry and ARM with Nvidia makes people uneasy but the emergence of RISC-V sees chips with a RISC-V core and an NPU (neural processing unit),” he said. “These are incredibly cheap, less than $10, with many out of China, and these will become ubiquitous,” he said. “I’m wondering if RISC-V will take over the world there.”
He is dismissive of a major programme at Leti working on spiking neural networks and analogue approaches such as resistive RAM (RRAM), but this might be expected from the inventor of the Convolutional Neural Network (CNN) and winner of the Turing Award for AI in 2018.
“The main problems that analogue implementations face is its very difficult to use hardware multiplexing with analogue neural nets,” he said. “When you do a convolution and reuse the hardware, you have to do hardware multiplexing and so you have to have a way to store the results and then you need analogue memory or ADC and DAC converters and that kills the entire idea. So unless we have cheap low power analogue memory that’s not going to work,” he said. “I’m doubtful, perhaps that will be memristor arrays or spintronics, but I’m somewhat sceptical.”
“Certainly edge AI is a super important topic,” he said. “In the next two to three years, it’s not going to be exotic technologies, it’s about reducing the power consumption as much as possible, pruning the neural net, optimising the weights, shutting down parts of the system that aren’t used,” said LeCun. “The target is AR devices with chips in the next two to three years with devices in the five years, and that’s coming,” he said.
Next: Pan-European edge AI technology platform
“Ten years from now is there going to be some breakthrough in spintronics or whatever that will allow analogue computation without hardware multiplexing?” he asked. “Can we come up with something like that that will reduce the size of the equipment so much for a single chip without data shuffling and without hardware multiplexing, that’s a big challenge,” he said.
“Companies are developing the technology of 1nm and 2nm for the next generation of chips and I strongly believe we can do things differently, with the sensor, the neural network and the controller for the future of the hardware,” said Emmanual Sabonnadiere, CEO of Leti. “We are trying to have national plans and have a piece of science in the political decisions. Edge AI is about stopping the data deluge and the privacy of data so people can own their own data.”
Leti also part of the Europe-wide neural network programme looking at a new platform for neural network chips.
- European research centres unite on neuromorphic memory
- Project looks at low-power chips for AI applications
“There is a new generation of technology to be investigated,” said Jean Rene Lequeypes, deputy CEO and CTO of CEA-Leti. “Now we are more than 2000 people working on what the next generation will be. Imec in Belgioum, Fraunhofer in Germany and Leti are working on one platform for edge AI and on top of that we are working in Inria in Grenoble to invent what we think the next support that Facebook and the big companies in Silicon Valley will need.”
The challenge is to integrate all the different elements without having to use the extreme UV lithography required at 5nm and below.
“We want to have the ultimate performance 1000TOPS/mW which is a very big challenge and how to work with the memories, the different technologies and how you integrate that without having to go to EUV,” said Lequeypes.
Related edge AI articles
- COMMITTEE PROCESSING IMPROVES IN-MEMORY NEURAL NETWORKS
- LETI SHOWS 9-CORE CHIPLET COMPUTER
- BLAIZE DETAILS ARCHITECTURE OF GSP CHIP
- PERCEIVE LAUNCHES EDGE INFERENCE PROCESSOR
Other articles on eeNews Europe
- Graphene electronics pilot line takes shape
- Opinion: Data centre AI drives $30bn AMD-Xilinx deal
- Microsoft and ARM partner for end-to-end IoT links to Azure
- Multicore GPU aims at 3nm for data centre designs
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
