[Interview] 'Semiconductor Legend' Jim Keller: "HBM is Too Expensive... Cost-Effective AI Solutions Are Key"

Exclusive Video Interview with Our Newspaper

AI Chiplet 'Quasar'... "Launching Next Year"
Continuous Investments from Korean Companies like Samsung and Hyundai Motor
"Recognition of Tenstorrent Technology"
Regarding HBM, "We Need to See if the Trend Continues"
Also Revealed Plans to Visit Korea in the Second Half of This Year

Jim Keller, CEO of Tenstorrent, who is regarded as a strong contender to topple the dominance of AI semiconductor leader Nvidia and change the landscape of the global AI chip competition, announced that the AI chiplet 'Quasar' developed by his company will be "released next year." The original plan was to launch it by the end of this year, but the timing was adjusted due to several technical issues.

Jim Keller, CEO of Tenstorrent, is responding to a video interview with Asia Economy in Santa Clara, USA. Photo by Zoom Capture Photo by Zoom Capture

CEO Keller, who attracted attention by securing investments from domestic companies such as Samsung Electronics and Hyundai Motor Company, attributed this to "the high evaluation of Tenstorrent's technological capabilities." A proponent of the open architecture RISC-V, he said, "It has tremendous advantages in innovation," expressing high expectations for its growth potential.

In a recent video interview with Asia Economy, CEO Keller explained, "We will release Quasar, which was originally scheduled for the end of this year, next year after supplementing technology development and addressing new design issues." The company added that they expect the release in the second half of next year.

Quasar is a so-called 'comprehensive semiconductor' being developed by CEO Keller and Tenstorrent in line with the recently popular 'AI chiplet' trend among companies. A chiplet refers to combining semiconductors with different functions into a single product. Quasar, designed for this chiplet approach, is 'packaged' to perform tasks that multiple semiconductors would otherwise share, all by itself. It is expected to reduce the burden on global companies that had to pay large costs to supply many semiconductors, making it a product boasting the best cost-effectiveness. The global semiconductor industry is paying close attention, suggesting that Quasar's launch could become a major variable in the AI chip market competition.

Interest is further heightened as Samsung Electronics is reportedly cooperating in the development and production of Quasar. Tenstorrent agreed last July to jointly conduct an AI chip R&D project with Samsung Electronics, and Quasar is said to be manufactured at Samsung Electronics' foundry plant in the U.S. using the 4nm (nanometer, one-billionth of a meter) 4th generation process (SF4X).

CEO Keller said, "Samsung is collaborating on the AI chip developed by Tenstorrent."

Tenstorrent has been accepting pre-orders for its AI chip 'Wormhole' since last month. When designing Wormhole, CEO Keller used graphics DRAM (GDDR6) instead of high-bandwidth memory (HBM), significantly reducing costs. He also said about Wormhole, "The advantage of being cheaper than Nvidia's products stands out, and I believe we have sufficiently achieved our initial goal." However, it showed limitations in performance, only reaching about 30% of Nvidia's products. Tenstorrent is developing Blackhole, the successor to Wormhole, which is currently being produced through Taiwan's TSMC. Blackhole aims to overcome these limitations and significantly reduce power consumption, positioning itself as a product that will surpass Nvidia's offerings in one go.

Below is a Q&A with CEO Keller.

Q. Hyundai Motor Group and Samsung Electronics are actively investing in Tenstorrent. What do you think is the background behind this?

A. Both Hyundai and Samsung chose Tenstorrent because of its technological capabilities. Especially in the AI processor sector, there are various reasons, but Tenstorrent possesses the technology needed for AI processors and chips in many components required for the new era of automotive technology. Samsung has the foundry capable of implementing such technology. There are likely other reasons as well. Also, the various relationships between Samsung and Hyundai probably made such investments possible. Both companies have all the technological capabilities needed before and after any manufacturing or process stages. Working with both companies is enjoyable.

Q. You seem to believe that high-bandwidth memory (HBM) is not greatly needed in AI chips. SK Hynix and Samsung Electronics are prominent in this market segment. Personally, how do you view the future of HBM?

A. HBM is a very cool and highly efficient technology. Many products are already on the market. However, I think it falls short in terms of cost efficiency. So, while it is indeed a great technology, it is too expensive to further develop. Currently, the market is trying to reduce HBM prices by using alternatives or adopting more cost-effective technologies. Ultimately, price is the biggest issue, and whether HBM can maintain cost efficiency while sustaining high technological performance remains to be seen.

Q. I heard that Quasar was to be manufactured and released using Samsung Electronics' 4nm process by the end of this year. Is there a possibility of further cooperation with Samsung Electronics on processes below 3nm? Also, are there any models planned for release after Quasar?

A. Quasar will not be released this year; the launch has been postponed to next year.

Q. Is there a particular reason for the delay to next year?

A. It is true that the original plan was to launch by the end of this year. However, due to technical and new design issues, it was postponed to next year.

Q. How are the initial sales of the Wormhole AI accelerator, which was introduced last month?

A. Wormhole was initially intended to attract more developers. It also has the advantage of being cheaper than Nvidia's products, so I believe we have achieved our initial goal.

Q. In past interviews, you mentioned that Nvidia cannot be considered a competitor. What kind of company do you hope Tenstorrent will become in the future?

A. Nvidia and Tenstorrent operate in different markets. Nvidia focuses on more expensive, high-performance AI products. We mainly develop AI designs in CPUs, low-class chips, and products that help companies easily develop AI solutions. Our goal is to create products needed to make AI chips easier to use in specific markets like automotive and to produce many products used in open-source technology.

Q. The U.S. has a well-established environment for fabless companies to grow, whereas Korean fabless companies are considered small and have a narrow market. As a leader of a startup, what advice would you give to Korean fabless companies?

A. The U.S. has a very healthy culture for startups to grow. No matter how many failures occur, people can learn from them. The U.S. environment is very positive for startups. Even if they do not succeed, they can take what they learned from failure and build value in the future. This allows startups to grow significantly. They receive opportunities to use as growth engines. That culture is substantial. Many startups gather in California and Texas. Several states provide foundations for startups to grow.

[Image source=Yonhap News]

Q. I saw an interview where you said AI will replace all software in the world. In a world changed by AI, what should humans learn?

A. I expect it will take some time for AI to replace all software. While AI is certainly a technology to be used in the future, there are still problems to solve to input more human capabilities and accomplish more tasks than now. Current AI requires people to input code or content directly to be usable. So, when developing applications, AI tends to work based on existing code rather than creating something entirely new. For now, humans seem to have more capabilities than AI, but I believe this issue will be resolved over time, and we need to watch how this progresses.

Q. Despite the technical excellence of the RISC-V architecture, why do you think it has not yet received significant attention in the market?

A. I think it is still an early stage. As you asked, more time is needed to gain greater attention. In fact, I think it is doing quite well now, and in industries using small or medium CPUs, this technology is already well utilized. Some countries like China are using higher-level products. The advantage of this technology is that anyone can freely modify it. This will be a great advantage in new innovation and development. I do not doubt the future of this technology.

Q. Do you have plans to visit Korea this year?

A. Of course. There are several events, and I am coordinating my schedule.

Interview by Choi Il-gwon, Head of Industrial IT Department

Compiled by Kim Hyung-min

Jim Keller, CEO of Tenstorrent, is widely recognized as a legend in the semiconductor design field. In the IT industry, he is often called a 'genius engineer.' Keller majored in electrical engineering at Pennsylvania State University and demonstrated his abilities through key positions at global companies after becoming an engineer.

The semiconductor chips he created greatly enhanced the reputation and status of those companies. At AMD, he worked as a chief architect (1998?1999) and vice president (2012?2015), creating the 'Hammer architecture (Athlon)' and 'Zen architecture (Ryzen)' series. These products helped AMD emerge as a formidable competitor threatening Intel, which was the market leader at the time.

At Apple (2008?2012), he developed the mobile application processors (AP) A4 and A5, known as the brains of smartphones. At Tesla (2016?2018), he created the autonomous driving chip 'Hardware (HW)3,' laying the foundation for full self-driving (FSD) technology. He also served as senior vice president at Intel from 2018 to 2020. After leaving Intel in June 2020, Keller became CTO of the Canadian AI chip startup Tenstorrent four months later. Since January last year, he has been CEO, leading the company overall.

Tenstorrent is a fabless startup developing AI chips. Founded in 2016, it is headquartered in Toronto, Canada, with branches in the U.S., U.K., Japan, India, Serbia, and South Korea. It excels in AI processor development and semiconductor design and recently has been providing solutions and IP licenses centered on RISC-V to companies.

Key Profile

Nationality: United States
Age: 66
Occupation: Engineer, CEO (Tenstorrent)
Career: AMD Chief Architect (1998?1999), SiByte Chief Architect (1999?2000), Broadcom Chief Architect (2000?2004), P.A. Semi Vice President (2004?2008), Apple Vice President (2008?2012), AMD Vice President (2012?2015), Tesla Vice President (2016?2018), Intel Senior Vice President (2018?2020), Tenstorrent CTO (2020?2023), Tenstorrent CEO (2023?)