Chip Startup Ceremorphic Seeks AI, HPC Edge With New 5nm Design

CEO Venkat Mattela says Ceremorphic is one of the few chip startups to have access to TSMC’s 5-nanometer node, but its real differentiation will from technologies it’s designed in house that will provide competitive advantages in efficiency, reliability and security for AI, high-performance computing and beyond.

Chip startup Ceremorphic is building a scalable, low-energy processor for AI, high-performance computing and other applications it said will stand out from competing chips with improved capabilities for efficiency, reliability and security.

The San Jose, Calif.-based startup, which recently came out of stealth mode, said it has an advantage over most other chip startups: access to TSMC’s 5-nanometer process node, which Ceremorphic will use for its Hierarchical Learning Processor when it enters production in 2024.

[Related: The 10 Hottest Semiconductor Companies To Watch In 2022]

Venkat Mattela, founder and CEO of Ceremorphic, told CRN in late January that access to an advanced manufacturing process like the 5nm node from TSMC — the Taiwanese chip foundry that makes chips for Apple, AMD and Nvidia, among others — is not trivial for a startup.

“This node access is provided based on the credibility of the company, expertise of the team that can handle such a complex node and also the ability to fund the project till the end,” he said.

That credibility, Mattela said, comes from the work he and his colleagues did at his previous semiconductor company, Redpine Signals, which sold its wireless assets to Silicon Labs in March 2020 for $308 million. From that exit, Mattela and Ceremorphic’s other founders pooled together a $50 million Series A funding round, which helped it expand from an initial 18 employees to a workforce of 150, many of whom work at the startup’s research and development hub in India.

“Unlike many of the very heavily funded startup companies, we are funded heavily because of our previous exit. Otherwise, TSMC would not give 5nm access,” he said.

Mattela said the access to TSMC’s 5nm node is significant because of the practical benefits associated with shrinking the transistors to such a miniscule level: “For example, in one square millimeter of silicon, we can create 40 million gates today in 5nm,” which he said would be equivalent to the processing power of a supercomputer from roughly 30 years ago.

But while Ceremorphic may be one of the few startups with 5nm access, it will be going up against larger semiconductor companies that are planning processors with the advanced node. That list of companies reportedly includes Nvidia, the dominant provider of processors for AI and HPC.

That’s why Ceremorphic’s real differentiation will come from the technologies it’s developed in house, Mattela said. They include a custom machine learning processor, a custom floating point unit and custom video engines that will go alongside an Arm M55 core in the Hierarchical Learning Processor, which Mattela described as an “ultra-low energy supercomputing chip.”

Some of these technologies are patented, including one for multi-threaded processing that is being applied to a RISC-V processor that will be used for what Mattela called “proxy processing.” The company has also built a custom PCIe 6.0 interface.

Mattela said it was important for his team to design as much of its own technology, including software, as possible rather than becoming too reliant on designs from companies like Arm.

“If I can buy an IP, obviously the bigger company has more opportunity. The only opportunity that I have is how do I differentiate some of the big functions,” he said.

Mattela said these custom technologies will give Ceremorphic an edge in three key areas: reliability, security and energy efficiency. And they will allow the startup to create chiplet designs for devices with different requirements, from virtual reality glasses to supercomputers.

For example, when it comes to energy efficiency, Mattela said Ceremorphic has designed its own algorithms that can reduce the size of workloads being processed by its chip. That, according to Mattela, will improve performance while also lowering energy requirements.

“That means if something takes 100 operations, I try to do it in 70 operations. So 70 operations means lower energy. Then I say I’m getting higher performance, and I’m getting low energy,” he said.

As for reliability, Mattela said, it’s an area that hasn’t received as much attention from other semiconductor companies, but he believes it is important for applications like AI training and advanced driver assistance systems. To solve for this, Ceremorphic’s architecture will use logic and machine learning “to predict faults and correct them using optimal hardware software approaches.”

But the real test for Ceremorphic lies ahead: Mattela said the startup will get its first test chips from TSMC in March, which will help it validate the performance it has seen in simulations. From there, the startup expects to provide samples to customers in 2023 ahead of a planned production in 2024.

“We are very careful in terms of when we do a silicon chip, because silicon development itself, independent of any technology, is a massive undertaking. So I always believe in creating differentiated technology in the foundation before we go and start doing the chip,” he said.