Eben Upton has hinted at a possible second in-house silicon design to come from the Raspberry Pi application-specific integrated circuit (ASIC) team, as a follow-up to its barnstormingly successful RP2040 microcontroller: Lightweight, low-power accelerators for edge AI and tinyML.
Raspberry Pi launched two firsts earlier this year: The Raspberry Pi Pico microcontroller development board, and the RP2040 microcontroller which drove it. Although the company has long designed its own single-board computers, a shift to microcontrollers came as a surprise — as did the news that the chip was a product of its emerging-from-stealth in-house ASIC design team.
With the RP2040 now out and appearing in third-party boards, questions have been raised about what the Raspberry Pi ASIC team is working on next — and co-founder Eben Upton has the answer: Lightweight accelerators for on-device ultra-low-power machine learning.
Speaking during a talk presented as part of the lead-up to the tinyML Summit 2021, Upton explained: "There's a Neal Stephenson book where somebody asks one of the characters, talking about laying cables under caves in the Philippines, 'do you plan to do this again,' and he says 'well, business people don't like to do things once because it messes up the spreadsheet.'"
"I think it is overwhelmingly likely that there will be some other piece of silicon like [RP2040] from Raspberry Pi. I think there's a big opportunity here: Because of its need to run efficiently on processors, the tinyML world has drive a real focus on good enough primitives. The interesting thing about this world for us is it's a very static world in terms of what the primitives look like, so there's a little bit of research interest at the moment in what you can build in the way of a slightly better implementation — something which probably has no more arithmetic throughput than a processor core, but doesn't have all of the scaffolding around it."
While the slide accompanying the talk suggested the design may take the form of lightweight accelerators capable of 4-8 multiply-accumulates (MACs) per clock cycle, Upton was hesitant to commit: "I think you might see something from us in the future on this," he suggests. "I wouldn't be surprised to see that find its way into a future piece of silicon. Raspberry Pi met tinyML, and discovered we weren't too bad at it."
The full presentation is now available on the tinyML YouTube channel.