In a market still filling with fledging silicon chips, Ceremorphic, Inc. has exited stealth and is telling the world about what it calls its patented new ThreadArch multi-thread processor technology that is intended to help improve new supercomputers.
Venkat Mattela, the company’s founder and CEO of Ceremorphic, calls his latest chip design a Hierarchical Learning Processor (HLP), even though several technology analysts said they recognize it as a system on a chip (SoC) design. The goal of the company is to design, benchmark and market a new kind of ultra-low-power AI training chip.
“What we are trying to solve is today – everybody knows how to do higher performance – you can buy an Nvidia machine,” Mattela told HPCwire. “Can we have the highest performance in a reliable way? Architecture is how we achieve it,” using multiple processors, a multiple logic design and mixing and matching it all. “What I am solving is reliable processing.”
The QS1 chip uses a 5nm design made by TSMC for Ceremorphic. It will mount into a 250 watt PCIe riser card and plug into a main system board.
The value proposition for Ceremorphic, according to the CEO, is that the HLP chip design will deliver better performance, lower power consumption and fewer failures compared to competitors.
“When [competitor’s systems] fail, nobody knows why did they fail,” said Mattela. “What I am saying is that the failures cannot be accepted when the number of processing requirements go up. I am coming with an architecture, and I make it fail less.”
So, why is he calling the new chip design a hierarchical learning processor?
“I do not want to do something which already exists,” said Mattela, explaining its new name.
Test Chip Expected on March 16
The company’s claims come even though none of the new chips have yet been built, evaluated or sampled, he conceded. Those tests will validate their designs and architectural features, he said.
The expected performance is theoretical, but it is based on patented technologies and calculations which promise such performance, he said.
Ceremorphic says its new silicon system aims to power next-generation applications such as AI model training, HPC, automotive processing, drug discovery, robotics, life sciences and metaverse processing. The chip architecture is designed to solve today’s high-performance computing problems in reliability, security and energy consumption to serve all performance-demanding market segments, according to the company.
Key features of the QS 1 include a custom machine learning processor running at 2GHz, a custom floating-point unit running at 2GHz, a patented multi-thread processing macro-architecture on a ThreadArch-based RISC-V processor for proxy processing (1GHz), custom video engines for metaverse processing (1GHz) along with M55 v1 core from Arm Ltd., and a custom-designed X16 PCIe 6. 0 / CXL 3.0 connectivity interface. It will also include Open AI framework software support with optimized compiler and application libraries. With these technologies, Ceremorphic is promising a soft error rate of (100,000)-1.
Mattela has more than 30 years of engineering and management experience in developing differentiated products and building successful businesses. Before founding Ceremorphic in April of 2020, he founded Redpine Signals, Inc., a wireless technology company that he sold to Silicon Labs, Inc. for $308 million. He holds a doctorate in electrical engineering from Indian Institute of Technology and is a graduate of Harvard Business School. Mattela holds more than 100 U.S. and international patents.
Ceremorphic was launched through a $50 million Series A investment from Mattela and his family, the CEO said. The company has about 150 full-time employees at its development facility in Hyderabad, India, and its business headquarters in Silicon Valley.
Analysts Weigh In
Rich Wawrzyniak, the principal market analyst for ASICs and SoCs at Semico Research Corp., said he sees what Ceremorphic is eyeing with their new silicon designs – a new architecture for AI training with high accuracy – but notes that they have not yet delivered any benchmarking data to prove their claims.
“I would suspect they have carried out many simulations of the architecture they are creating, so [they probably] have some data to back up what they are saying,” he said. The test silicon run is coming up in March and the company is planning customer samples in 2023 and chip production in 2024, he added. “Many things can change between now and production, and I expect them to refine their numbers between now and then. Because of this, they probably do not want to give out too much info on the simulations since they will probably change quite a bit. You want to be as accurate as possible when it comes to performance data.”
Ceremorphic’s performance aim is for close to or at 100 percent accuracy, “and this would be a big advance if they can do it,” said Wawrzyniak.
Under Ceremorphic’s designs, the chips would use the CPU power of the host system to provide the computational power and then Ceremorphic would provide acceleration to the training part of the operations, he explained based on his knowledge of the systems. “Venkat seemed to say that their architecture would provide protection against attacks by quantum-enabled systems, but he did not elaborate on how this would be accomplished.”
Overall, Wawrzyniak said the company’s planned chips seem workable.
“I would say they have a fair chance of success with this approach,” he said. “It is much different from what others are doing with their AI architectures.”
Another analyst, Linley Gwennap of The Linley Group, told EnterpriseAI that Mattela “has a good track record. What they have disclosed so far does not appear revolutionary. It is an SoC optimized for AI. Many startups and established companies are competing in this space and have a head start on Ceremorphic.”
In addition, Ceremorphic’s claim that it has developed a completely new architecture is the same claim of every other AI startup as well, said Gwennap. “Since Ceremorphic did not disclose any specifications for performance or power consumption, it is impossible to evaluate the company’s competitiveness. Once we get closer to first silicon, I assume it will release more details.”
Karl Freund, the founder and principal analyst of Cambrian Research, agreed.
“It looks like ‘Yet Another ML Processor (YAML)’ to me,” said Freund. “[It is an] all-custom SoC, but interestingly combines Arm and RISC-V with their own GEMM. The market is crowded, and they are late to the party, but their board of directors and advisors from leading universities are impressive. I cannot wait to see the results.”