basement: Legendary chip designer builds team in B’luru | News from Bengaluru

Bengaluru: Legendary chip designer Jim Keller is building a team in India while working to develop a specialized chip to run AI programs.
In an exclusive interaction with TOI, Keller, who is CTO at Tenstorrent, said about nine people in India have accepted their offers. “We’re building an AI processor, we’re building a software stack on top of it, we’re building a general-purpose CPU based on RISC-V (architecture), we’re building infrastructure software, and we’re looking for good people who can do that and want to work in a startup and want to do new things,” he said. The company will have its office in Bengaluru.
There is really great talent in India, he said. “I have worked with engineering teams in Bengaluru and Hyderabad. I like to go where the energy is,” he said.
Tenstorrent was founded by a team that included Ljubisa Bajic, who previously worked at Nvidia and AMD. At AMD, Keller and Bajic met for the first time. When Bajic came up with the idea for Tenstorrent, Keller financed it with seed capital. In January of last year, Bajic lured Keller into joining Tenstorrent. Four months later, the company received $200 million in funding worth $1 billion.
Keller’s career has spanned a number of notable companies – Digital Equipment (which was acquired by Compaq), Broadcom, Palo Alto Semiconductor (PA Semiconductor, which was acquired by Apple), Apple, AMD, Tesla and Intel. At Apple, Keller worked on the earliest in-house chips for iPhones and iPads. At AMD, he helped build the Zen processor, which had a major impact on AMD’s future. At Tesla he designed the latest self-driving chip called Hardware 3. Hardware 1 was a Mobileye solution, Hardware 2 was an Nvidia solution, but Hardware 3 was built in-house. “It’s now four years in production and it’s still the best autonomous driving solution for the money I’ve seen. We beat that out of the park,” says Keller. The work involved a partnership with Samsung who performed the backend/physical design – the process of turning a design into manufacturable geometries. A large part of the physical design team at Samsung has been in India, says Keller.
Today, trying out an AI processor, he finds himself in a crowded room. Nvidia and AMD are adapting their GPUs to develop AI chips. Intel is adding AI accelerators to its CPUs and is also building standalone AI processors through Habana, an Israeli company that acquired it. And there are others like Google, Graphcore, SambaNova, and Groq that are also building AI processors. The idea behind this is that the requirements for AI are unique and chips need to be optimized for deep neural network training and for the use of inference in production environments. You not only have to analyze huge amounts of data, but also a variety of data – images, videos, sentences and even abstract concepts.
According to Keller, AI programs are written in PyTorch and TensorFlow, and Tenstorrent is striving to build a standalone chip that runs AI programs as written in PyTorch and TensorFlow, minimizing the hand-hold of an Intel or AMD processor. GPUs, he says, are designed to run graphics programs on pixels. CPUs were designed to run assembly languages, and then C, and then high-level languages. “We’re trying to build a computer that actually runs AI programs like those written in PyTorch and TensorFlow. It seems obvious when you say you should build a computer to run a program the way people write it, but it’s a tough problem. And that’s the problem we’re working to solve,” he says.

Comments are closed.