Netronome is helping to lead a new open group that aims to disaggregate various functions of silicon chips as a way to overcome the slowing down of Moore’s Law. The group is called the Open Domain-Specific Accelerator (ODSA) workgroup, and it’s being led by Netronome along with six other silicon companies.
Engineers are scrambling to figure out ways to compensate for the slowing down of Moore’s Law. It’s a big dilemma because the rate at which server performance improves is slowing down. And at the same time, data growth is exploding.
One way network engineers have coped with the tension is to offload some functions from servers onto network interface cards (NICs).
Netronome was one of the first companies to develop a smartNIC, which it’s been selling for about three years. But in its dealings with customers, Netronome has been getting requests for domain-specific accelerators.
Domain-specific accelerators are programmable chips that are tuned for specific kinds of workloads. Bapi Vinnakota, director of silicon architecture program management at Netronome, said typically if you have a fully programmable CPU on a server, you can run any kind of code on it. “The catch is: will it run efficiently?” he said.
He explained that domain-specific accelerators run special classes of workloads. “The most well-known one is Google, which spent resources on accelerating machine learning workloads,” said Vinnakota. “They offloaded the host and built a special piece of silicon to do a wide range of neural code efficiently.”
The idea with domain-specific accelerators is: rather than building a monolithic chip with a full range of capabilities, decompose the chip into multiple “chiplets.”
The Open Domain-Specific Accelerator
Netronome formed the ODSA group in October 2018, along with six other companies: Achronix; Global Foundries; Kandou; NXP; Sarcina; and SiFive. The group plans to develop an open architecture and related specifications for developing chiplets for domain-specific acceleration.
One of the reasons for creating the ODSA group is that silicon development is expensive.
Vinnakota said ODSA faces “two countervailing pressures.” The first is to try and beat Moore’s Law by building domain-specific accelerators. This will be costly. At the same time, a domain-specific accelerator targets a smaller market than a CPU.
The members of the ODSA group have a strategy for building their open, domain-specific accelerators. First, each member of the group will contribute its particular technology expertise. For example, Netronome brings its intellectual property from its Open Network Flow Processor (NFP) architecture.
Secondly, the group will develop technology to connect its domain-specific accelerator chiplets to other ASICS and programmable chips. That way it doesn’t have to re-invent the wheel for the parts of the chip that already work fine. Hence, the disaggregation aspect of the chip.
“You build a chiplet with new IP that you have and you use it to talk to IP that you need,” said Vinnakota. “The tech challenge is how to make them talk to one another almost as efficiently.”
“You need these chiplets to share an architectural interface and to believe they’re all part of the same chip,” he said. “Until now that’s been entirely proprietary. We’re going to make that architectural interface completely open. If you can make it open, you can assemble chips from multiple vendors.”
Brian Sparks, VP of marketing at Netronome, said the ODSA group has released its initial white paper with prototype tech models as well as business models. It’s planning its first workshop at the end of this month. Eventually, it hopes to join a larger open source organization.