Taalas Inc., a startup that plans to sell chips tailor-made for specific artificial intelligence models, has raised $50 million to support its commercialization effort.

The company announced the raise on Tuesday. The funding was provided by Quiet Capital and Pierre Lamond, a prominent venture capitalist who backed several of the semiconductor industry’s early players and whose chip bona fides goes back to Fairchild Semiconductor and National Semiconductor in the 1960s.

Taalas is led by Chief Executive Officer Ljubisa Bajic, who previously launched AI chip startup Tenstorrent Inc. in 2016. The latter company received a $1 billion valuation after its most recent $200 million funding round. Bajic founded Taalas last August with Drago Ignjatovic and Lejla Bajic, who held engineering leadership roles at Tenstorrent.

Practically all AI chips include optimizations designed to speed up matrix multiplications, the mathematical operations that neural networks use to process data. Some processors feature additional optimizations geared towards specific AI use cases. Nvidia Corp.’s latest H200 graphics card, for example, ships with a large amount of high-speed memory to accelerate language models.

Taalas plans to take the concept a step further. According to The Information, the company is working on processors that won’t simply be optimized for AI, but will be built with the requirements of a specific neural network in mind. The company hopes that this approach will make its chips significantly faster than today’s graphics cards.

“Commoditizing AI requires a 1000x improvement in computational power and efficiency, a goal that is unattainable via the current incremental approaches,” Bajic said. “The path forward is to realize that we should not be simulating intelligence on general purpose computers, but casting intelligence directly into silicon.”

Developing a custom processor can take years and hundreds of millions of dollars in some cases. As a result, Taalas’ plan to create multiple chips that will each be optimized for a different AI algorithm is likely to involve significant technical challenges. To address those challenges, the company is developing an automated engineering workflow that it says will accelerate its semiconductor design efforts.

Taalas says one of its processors will contain enough memory to hold an “entire large AI model.” Running a model entirely on-chip removes the need for external RAM, which means data doesn’t have to travel between the RAM module and the chip. Reducing data travel, in turn, speeds up processing.

Taalas plans to deliver the tape-out of its first product, a processor optimized for large language models, in the third quarter. A tape-out is the version of a chip design that is sent to the fab for manufacturing. Taalas will start shipping the processor to customers in early 2025. 

Photo: Unsplash