A chip expert said that with the help of open source technology, the development of artificial intelligence chips will become easier and cheaper, "as long as the hardware chips are truly programmable, people can use their creativity to build new solutions."
Jim Keller, an American microprocessor engineer who has worked at AMD, Apple, and Tesla, told Anadolu Agency, "AI processors are actually simpler than people think."
"People want you to believe that developing an AI processor requires $100 billion, but you don't," he said on the sidelines of GITEX Global in Dubai, one of the world's leading technology and AI events.
Keller said his company, Tenstorrent, develops open-source technology from AI processors to general-purpose processors, and they have also open-sourced their AI compiler, making it publicly available.
He noted that open-source chips are less expensive and their architectures are more accessible.
"The current models are very good and are constantly improving," he said. "I read an article every day saying something has reached its limit—and we're not even close to it. The demand for it is really, really huge; so, I don't know, the next five years are going to be very interesting."
Keller said that restrictions on the chip industry imposed by some countries won't be effective in the long run, "because some of the technology is really open and people have published a lot of results," he said.
"It turns out that when you restrict an area, you just force it to evolve, and so far, it's just meant they're playing catch-up," he said. "I hope we're moving towards a more open, less restrictive world."
"Every company is different, right, and their cultures are different," he said, referring to his time at AMD, Apple, and Tesla.
"The most effective way is when the team is working towards a really good goal, like a great product, and then I really like when engineers can own their work and be truly inspired by what they do every day and the people they work with."
"I want to create a really open environment where everyone can learn, build things, and control their own destiny," he added.
Jim Keller on AI, RISC-V, and More
Legendary CPU designer Jim Keller, formerly of Apple, Tesla, and AMD, will take over as CEO of AI chip company Tenstorrent in early 2023 after two years as CTO. Keller has been a staunch supporter of RISC-V in recent years, and this burgeoning open-source ISA was a key topic of discussion in an exclusive video interview with EE Times.
"I believe that within the next five to 10 years, RISC-V will take over all data centers," Keller said, adding that this is particularly true for scientific computing and high-performance computing (HPC). He suggested that supercomputing's dominance could happen even faster.
Keller is a strong believer in open source hardware and software; Tenstorrent plans to open source its AI software stack soon.
"We wanted to do this last year, but we weren't ready—our software stack was too messy, and we needed to partition it in the right way," he said. "We aspire to be a successful open source software company."
In a surprising move for a data center chip company, Tenstorrent recently licensed its Tensix AI accelerator core IP and Ascalon CPU core IP to LG Electronics. The South Korean consumer electronics giant plans to use Tensix IP for embedded edge computing, such as smart TVs and automotive chips. The two companies also plan to collaborate on future generations of RISC-V CPUs, AI accelerators, and video codec IP and/or chipsets. (LG spun off its own AI IP division in 2020.)
Keller described customer interest in Tenstorrent's IP cores as a "pleasant surprise" and said the company is still in business exploration mode.
"To be honest, we're just getting started...it's still small, and there are a lot of technical issues to work out," he said in our exclusive video interview. "As a technologist, I'm very happy to talk to smart people. So, if a smart person says, 'I want open-source access to your hardware so I can program it the way I want,' how can I say no?"
Keller's view of the edge AI market is that, to date, alternative IP offerings have been too concentrated and too difficult to program.
"We've gone through five transformations in the last two years," he said. "Is it image or language? Is it inference or training? Is it big models or small models? Is it generative or non-generative? Everyone who targeted a particular piece of intellectual property, the next model wouldn't work. We were founded on the premise that AI would advance so rapidly that the lines between inference and training, language and image, would blur and they would switch back and forth."
Keller's example, StableDiffusion, is part image model, part language model, with a backward pass in the training pipeline—in his words, "it looks like the kitchen sink—which IP is it running on?"
Source: Content compiled from tribune
