EDA industry veteran Wally Rhines has joined many-core accelerator chip company Cornami as CEO.
Rhines, who retains his position as CEO Emeritus at Mentor Graphics following almost 25 years as CEO, is a high-profile hire for Cornami. The Campbell, CA, company is currently a roughly 60-person startup seeking additional funding to finance production of an accelerator chip expected to tape out in 3-4 months’ time.
Cornami started out developing software that helped applications take full advantage of multi-core processors by using them in the most efficient way. This concept was developed into a hardware design for a reconfigurable computing fabric which can scale from multi- to many-core designs — thousands on a chip or millions in a system — plus a proprietary high-speed network-on-chip for communication.
Cornami is engaging with customers about accelerating AI workloads, particularly in autonomous driving where the chip can be configured as a full processor including AI acceleration (as opposed to an AI co-processor) where it offers performance and latency advantages.
However, the company’s future may lie in an intriguing related field where there are currently no competing solutions.
Getting the job
Speaking to EE Times by video link, Rhines explained that a meeting with Gordon Campbell, another well-known industry veteran who is a co-founder of Cornami, piqued his interest about the company.
“I had no intention of becoming a CEO again, but I was winding down my amount of time with Mentor,” Rhines said. “I’ve seen the spectrum [of AI chip startups] and it’s a bit of a crowded field, although a lot of them are somewhat specialized and in different niches… but valuations have been quite good. If you look at some of the leaders, they’re averaging a little over $2 billion valuation pre-revenue — that’s pretty heady territory.”
Discussions with Cornami’s team revealed impressive high-performance, low-power figures compared to existing accelerator offerings on the market. Rhines admits to a little skepticism at this point, based on his experience of how “better, faster, cheaper” would still be a tough sell in a crowded field that has already attracted significant investment. However, an aside at the end of Cornami’s presentation mentioned fully homomorphic encryption (FHE).
Separate from his discussions with Cornami, Rhines had been contracted by Darpa to find out whether the semiconductor industry was capable of developing a chip to accelerate this tech, a key part of the US DoD’s strategy of “securing the data, not the data center.” Rhines had investigated, and determined that no such chip was imminent. And here was Cornami co-founder and CTO Paul Master, in an off-hand comment, making the claim the company could make FHE possible.
“When he said we can do FHE, I said: come on! I know you can’t do it, because I’ve just surveyed all the leading semiconductor companies, and all of them told me this is 10-20 years away,” Rhines said. “[Those companies] patiently explained to me that this would require a million times the performance of the best Intel Xeon or Nvidia GPU servers of today. So you can imagine my skepticism when Paul said, yeah, we can do it. And not only that, we can do it essentially in real time, and we can do it for about the same cost as an Nvidia server.”
Cornami’s partners up to that point had been in the automotive, 5G and robotics markets and were interested in accelerating machine learning applications in one form or another. But Rhines sees Cornami’s future in the acceleration of FHE.
“If you can do all these things, why not do the one where there’s no competition and the one that has almost unlimited value placed upon it by people who want to secure their data?” Rhines said.
While it was too late for Darpa, which according to Rhines decided to fund a 5-year program to develop its own FHE chip, opportunities may lie in the financial sector and the cloud, where there is pressure to keep data secure.
“FinTech people can move quickly if they believe there’s a solution,” said Rhines. “They said: just ship us a board or a chip, and we’ll buy whatever you can make, and we’ll set up our evaluation platform and then we’ll push it into the cloud.”
This agility is much more attractive than some of the markets Cornami is already pursuing; while automotive will evolve to be a big market for AI acceleration, the design cycle is notoriously long.
Many, many cores
Cornami started as a software company, developing software to take advantage of efficiencies provided by multi-core processors. This software allows linear performance improvement with the number of cores and number of chips. The company developed a hardware compute fabric around the same idea — a reconfigurable systolic array where data flows through the chip without having to be buffered in caches or written off-chip, hence its suitability for AI. Cornami cores are independently programmable and perform mathematical operation on data of any precision level. The Cornami chip will use thousands of cores, scalable to millions in multi-chip systems.
Fully homomorphic encryption is a method for encrypting data which allows any mathematical operations to be run on encrypted data without decryption (similar to plain homomorphic encryption schemes already used commercially, but these typically allow adding numbers together but nothing more).
FHE entails “thousands of FFTs sequentially, and 500 order polynomials with coefficients that are double precision floating point,” according to Rhines — in other words, it’s so incredibly complex that regular computers can take minutes to do a single bit of a calculation.
Cornami chips can accelerate FHE to usable levels, enabling the technology to fulfil its promise of performing AI on encrypted data in the cloud. Sensitive data such as medical or financial information, which previously wasn’t able to be processed in the cloud because of privacy, would become very valuable.
The reconfigurable aspect of Cornami’s design is particularly suited to FHE as a number of emerging libraries are still competing for adoption. This includes Microsoft’s simple encrypted arithmetic library (SEAL), IBM’s HELib, the open-source Palisade project, and others.
Cornami has had an FPGA-based emulation of its chip available for a year and a half. Full verification will take another 3-4 months, Rhines said, with silicon expected in the first half of 2021.The company is actively fundraising at the moment, having raised just under $30 million to date.
Related articles:
- ‘Holy Grail’ of Encryption Could Be a Game-Changer for AI
- Rhines: Deep Learning Will Drive Next Wave of Chip Growth
- Infineon’s $10B gambit, Wally Rhines at DAC, IEEE ‘Reviewgate’
Leave a Reply
You must Register or Login to post a comment.