India’s answer to superfast optical AI computing

It is becoming inefficient to send mushrooming volumes of increasingly high bandwidth data to centres hundreds of kilometres away for AI engines to crunch for applications like video analytics to detect unsafe practices in covid-19 times or to monitor remote factories. It makes more sense to process the data at the edge, that is, locally. But that requires small high-power devices that will consume low power.

Hyderabad-based LightspeedAI Labs is building optoelectronics processors to solve this problem. It uses optical interconnects instead of the traditional chip architectures that are less suited for edge AI. Using light to transmit information has already revolutionized broadband with fibre optics. But making compact devices that can compute at the speed of light has been the holy grail.

“It’s like compressing the performance of a desktop into a smartphone," says Rohin Y., chief executive and co-founder of LighspeedAI. “There are only a few in the world who can do this. That’s where the expertise and experience of our chief technology officer and co-founder P.V. Ramana comes in,"

Ramana has over 30 years of experience in optoelectronics and integrated circuits. Starting out as a DRDO scientist under Dr Abdul Kalam, he went on to work for Agilent, Avago and Philips before becoming a consultant for several hardware MNCs. He began with LightspeedAI as a consultant last year when the startup was founded, then came fully on board a few months ago.

Rohin, who started the company, is a BITS Pilani electronics engineer with diverse experience in industry, academia and entrepreneurship. His earlier startup was MyTi Technetronics, an R&D venture in electronics and design.

Coming together

LightspeedAI was in stealth mode until this month when it announced an undisclosed amount of funding from YourNest and growX Ventures. Incubated at the Semiconductor Fabless Accelerator Lab, it has also been selected for the Bosch DNANxt programme as well as to the Applied Materials Astra 2020 cohort.

The startup’s innovation comes from its frugal approach to building a world-class product. “For example, if I have to develop a new processor, it will cost half a million dollars. Instead, we work with existing processors but come up with alternative mechanisms (like optoelectronic connectivity)," says Ramana.

With the designs in place, the next stage is fabrication for a minimum viable product, expected in six months. Hardware products typically have a two- to three-year development cycle, so the year-old LightspeedAI appears to be much faster off the blocks.

“We don’t develop any of the components but integrate them into a viable product in an innovative way. Because we use proven components from the market, our time to market is much shorter compared to what others do," adds Ramana.

Research on optoelectronics has been going on for over two decades at premier institutions like Massachusetts Institute of Technology, which has a Microphotonics Center. Google Ventures has invested in a three-year-old US startup, Lightmatter, whose founder Nicholas Harris is an MIT alumnus. Lightmatter recently unveiled an optoelectronic test chip in which clusters of processors exchange data at high speed using optical signals.

China’s Baidu Ventures has invested in a rival Boston-based startup, Lightelligence, whose founder Yichen Chen was also a PhD student at MIT’s Microphotonics Center and even co-authored a paper with Harris, which forms the basis for the two startups.

The differentiator for the Indian startup, says Rohin, comes from its frugal approach. LightspeedAI focuses on the communication between chips, whereas others have been mostly going after the computing side of optoelectronics. “Our technologies and the purpose are quite different. They are focused on ‘computing’ with light; we are doing ‘communication’ with light," says Rohin, when asked specifically about Lightmatter and Lightelligence.

Secondly, there are several market segments in this nascent space, from the requirements of massive data centres for AI to many use cases of edge computing that LightspeedAI hopes to enable with its low-cost compact designs.

“Our system, a small ‘dabba’, can sit near a 5G antenna or a factory floor, processing data from multiple video cameras, or a company’s on-premise data centre for AI. This is modular and you reconfigure it for different applications," says Rohin. “Our computing is generic, that is, it can do any application. Our rivals’ chip computing is limited to specific applications."

Standing out

The startup has filed four provisional patents around its modular design and reconfigurability.

Essentially, it takes away the need for inefficiently sending data back and forth from data centres by enabling high-power computing at the edge. There, LightspeedAI has a direct competitor in San Francisco-based Ayar Labs, backed by Intel and DARPA (US Defense Advanced Research Projects Agency). It announced a $35 million funding round this month, taking its total funding to $65 million.

That makes it a David vs Goliath fight, admits Ramana, but the Hyderabad-based startup’s value proposition may lie in what’s done best in Indian deep tech: frugal engineering. LightspeedAI’s understanding of emerging industry requirements comes from Ramana’s experience of working for Singapore’s Institute of Microelectronics (IME). “Singapore’s research institutions are tailored to meet industry requirements. I had to get 30% of my budget from industry, so I had to go and look for problems and solve them," says Ramana. “IME also had an excellent scheme where the government subsidized researchers to go and work with small and medium enterprises for low fees. So, I was working with a lot of SMEs who cope with frugal budgets and need to deliver in a very short period of time."

The technology can be a differentiator because the way in which optical connectivity is being done between chips is different at Ayar Labs and LightspeedAI, although it doesn’t want to disclose its tech specs at this point. For an investor in such a space, the biggest risk is in the technology, when it’s hard to predict which solution will prevail in the market.

“But that’s also where the most value can be created for a deep tech startup, if they are able to demonstrate their technology works," says Manish Gupta, principal, investing, growX Ventures.

Apart from funding, where larger follow-up rounds will have to come from international VCs, LightspeedAI will also face significant go-to-market challenges. Its main target customers are abroad, because it’s the developed markets that have an appetite for new, disruptive solutions.

Malavika Velayanikal is a Consulting Editor with Mint.