Our aim is to make AI inference chips' more energy-efficient: Positron CEO

Mitesh Agrawal, CEO of US-based chip design startup Positron, claims their chips are three-and-a-half times better in terms of power usage and inference workloads

US-based chip design startup Positron’s mission is to make running artificial intelligence (AI) workloads and AI inference models as efficient and affordable as possible in terms of performance per dollar per watt, the company’s chief executive officer, Mitesh Agrawal, said.

 

Speaking to Business Standard virtually, Agrawal said the company’s proprietary chips, server racks, and hardware solutions have shown an up to 66 per cent improvement in energy consumption over other market leaders, especially in the AI inference workload segment.

 

AI inference is the ability of trained AI and large language models (LLMs) to identify patterns and draw conclusions from completely new data. While running inferences in LLMs uses nearly the same compute power as training of AI and LLM models, the electricity consumption can be significantly higher, especially for large-sized LLMs.

 

“The number one discourse, whether in the United States, India or any other Asian country, is that we have limited power, but we have to build lots of new data centres to support the workloads of all these new AI applications. For every given chip, if we can improve the performance 2-3 times, then we can generate that many more tokens and run as many more applications overall,” Agrawal said.

 

Positron has so far raised $75 million from investors, including Valour Equity Partners, Atreides Management, and DFJ Growth, in the US. A majority of this is currently invested in the research and development of newer chips that are even more energy-efficient than those being shipped by Positron right now, Agrawal said.

 

“The majority of the money right now goes into R&D as well as taping out the chips. Some of the money is allocated for our headcount, but the majority is invested in taping out. For tape-out, we are going to work with TSMC (Taiwan Semiconductor Manufacturing Company Limited),” Agrawal said. 'Taping out' refers to the conclusion of a chip's design process, after which it is considered ready to be sent for fabrication.

 

Although most chip design and manufacturing startups and companies typically take 4-5 years to tape out their first product, Positron was able to tape out and commercially ship its first product, Atlas, within 18 months of its launch in 2023, he said.

 

“We ran production workloads for the clients to show them that our product was 3.5x better than rival chips in terms of power usage and inference workloads. That is how we got our meetings and the clients took us seriously,” Agrawal said.

 

The startup, which currently operates with a lean team of 30-odd people, all of whom are based in the US, aims to expand to up to 50 people by the end of this year, he said.