Advanced Computing in the Age of AI | Monday, July 15, 2024

NASA to Upgrade Computing in Space with RISC-V 

NASA's Perseverance Mars rover. Image credit: NASA/JP:

NASA realizes its probes and computers in space are outdated, and a massive upgrade is needed as the agency starts unlocking the mysteries of unexplored frontiers.

NASA is now rebooting its spaceflight computing hardware, starting with RISC-V chips made by Microchip and SiFive. The Mars rover called Perseverance was sent into space last year with an unimpressive 20-year-old PowerPC chip.

NASA has said the new Microchip product will provide "at least 100 times the computational capacity of current spaceflight computers."

Computers deployed in space need to be radiation hardened and withstand lower temperatures. At the same time, the systems need to be power efficient with the limited amount of power available. The chip will have SiFive’s X280 cores, which include vector processors for AI inference computing.

NASA is trying to free itself of the computing constraints in its previous probes, and are trying to figure out how they can use AI in the newer systems, said Chris Jones, vice president of product at SiFive

NASA has been internally developing chips for space computers based on RISC-V, which is a free-to-license instruction set architecture. RISC-V is often referred to as the Linux of chips as the architecture can be customized with proprietary extensions. NASA is also developing internal RISC-V cores for in-orbit computing.

NASA is one of many academia, research and defense organizations backing RISC-V.

"They want to be able to do some amount of AI inference on the probe itself, instead of having to send all that data back here," Jones said, adding "they're trying to get as [much performance] as they can get through the given power budget to be able to do these sorts of navigation type functionalities."

To be sure, NASA hasn’t elaborate on the types of space devices getting the RISC-V chips.

SiFive and Microchip competed with the "usual suspects" for the NASA contract. The chip industry regulars include x86 chip makers Intel and AMD. NASA has deployed HPE’s Spaceborne computer in the International Space Station.
"Given the uniqueness of the architecture we put forth, and what NASA thought was more legs in RISC-V going forward, we were able to win this along with Microchip," Jones said.

SiFive and Microchip had to meet a set of benchmarks called Spacebench.

"We had to meet and really surpass the competition, which the X280 did, because it was well suited for that," Jones said.

NASA also wants to see a lot of freely available open software, which includes math libraries, linear algebra libraries and trig functions that people can use to build applications for the NASA spaceflight program.

"If you look at the amount of activity and interest and rapid rate of expansion of the RISC-V Foundation, you know, it's not going anywhere. You could have made an argument in 2016 that this was a toy... but six years later, it is a movement that has a ton of momentum," Jones said.

RISC-V is also seeing a groundswell in software development and an open-source community that NASA could tap into, Jones said.

NASA also chose RISC-V as it has big software ambitions for next-generation space exploration to the moon.

"We're also having this guidance from NASA regarding what their priorities are for us to put in the market to give people a heads up on how to program stuff," Jones said.

NASA also gave the SiFive and Microchip teams a suite of representative code examples that they want to run on their computers. Some examples included sensing, navigation and other technologies found in the latest manual and autonomous cars.

The architecture disclosed by Microchip includes 12 X280 chips hooked together in a very configurable way. That helps NASA dial up performance as they need based on what devices are doing.

SiFive augmented a RISC-V design with extensions such as matrix multiplication instructions for data types and instruction mapped interfaces to allow for easy connectivity to hardwired accelerators.

The chip's core feature is vector processing, which is back in fashion with AI. Vector processors started off in mainframes, but went out of vogue after CPUs took over.

"The actual silicon will come from Microchip so I don't know what their schedules are," Jones said, adding that he anticipated the hardware to come out in the next few years.