How an AI chip can bring in more reliability in supercomputing

If your phone or laptop hangs, you quickly go for a reboot and the device is ready for use again in the next 4-5 minutes. It’s not the same in case of a datacentre or a high-performance computing processes involved in drug discovery where you are dealing with humongous data driven by artificial intelligence and machine algorithms in a super computer. Any error could hamper the process and any rebooting could mean pushing the project schedule by a few weeks or even months.

AI chip QS1

Ceremorphic, a two-year-old US semiconductor start-up with its research and development in Hyderabad, has designed an artificial intelligence supercomputing chip, named QS1, that demonstrates a ‘reliable computing performance’ working at a lower energy profile.

Ceremorphic’s Founder-Chief Executive Officer (CEO) Venkat Mattela says the chip provides the performance required for AI model training, metaverse processing, automotive processing and drug discovery.

It addresses the problems of reliability, security and energy consumption confronted in high-performance computing (HPC) today. Ceremorphic is using TSMC’s (Taiwanese Semiconductor Manufacturing Company) 5-nanometre process node, making it one of the earliest users of the technology.

Also read

IIT Madras, US university develop AI-powered algorithms to enhance 3D effects in phone videos

Ceremorphic, which has a repository of 100 patents to its credit, is aiming to make the power of supercomputing accessible, affordable and mobile. It has recently opened a R&D centre in Hyderabad with 150 engineers.

Origin story

Mattela founded Ceremorphic with a small team of 18 engineers that he retained from the Redpine Signals asset sale in March 2020. (US-based Silicon Labs acquired Redpine Signals’ WiFi and Bluetooth business and development centre in Hyderabad in a $308 million all-cash deal).

“The technology we developed then was a highly differentiated wireless product in energy efficiency. It is 26 times better than the best in the industry,” he said. As the team had already started working on the ‘reliable performance computing’ chip well before the startup’s inception, it had got a head start in developing the chip.

“We are going to test the chip in September 2022, before going for full mask chip next year. We are expecting to go for production the following year,” Venkat told BusinessLine. The San Jose (the US) based start-up has put in $50 million in Series-A so far, raised from family and friends of Venkat, and other previous investors and employees. “We will raise over $100 million next year to fund the growth plans,” he said.

“Everybody knows how to do supercomputing. For the first time, we are working on reliable performance computing. A computer doesn’t fail too often but when it does and if there are too many of them, it will be very difficult to find out where the failure has occurred. If you can’t find it out, it will be difficult to get it fixed,” he said.

“We need an affordable supercomputer required to let everybody write programmes, making the AI/ML programming pervasive across all markets. Today, it is not something very affordable. It’s very, very expensive, and it is huge,” he points out. “What I’m trying to do is if I make the supercomputer in a tiny form factor, they can attach it to a laptop and get access to super computing.”

Difficult programme

He says the development of a chip is an arduous task. “The development is happening for the last two years and we still aren’t there yet. I don’t have anything to show after five years of R&D and two years of development,” he said. “However, a fully validated technology portfolio and a highly capable and technical team is built to execute this programme in the next 18 months,” he said.