In early 2016, an engineer approached Tesla CEO Elon Musk with a bold proposal: “Give me 18 months, and I’ll develop self-driving chips [for Tesla].” Musk, who was committed to advancing self-driving technologies, paused to consider the offer. “Alright,” he finally replied, as he accepted the engineer’s proposal. “But prepare a backup plan in case you fail.”
Exactly 18 months later, the engineer delivered an AI self-driving chip called HW3. It was a pivotal moment that laid the foundation for Tesla’s fully self-driving (FSD) technology. The engineer who made this possible was Jim Keller, the legendary chip architect best known for his work at Tesla, Apple and AMD. Keller is the mastermind behind Tesla’s original self-driving chip and was the lead architect of the Athlon and Ryzen processors that made AMD what it is today. At Apple, he was involved in designing the A4 and A5 mobile application processors, which serve as the iPhone’s brain.
ChosunBiz recently had the opportunity to talk to Keller via Zoom. He is currently the CEO of Tenstorrent, a Canadian AI chip startup, where he oversees chip design. Tenstorrent unveiled the Grayskull processor, which boasts a performance of up to 315 trillion operations per second (TOPS) on INT8 (8-bit integer). Despite its high performance, Grayskull is more power-efficient than other AI chips in the industry, operating at 200W. “Tenstorrent’s competitive advantages are high power efficiency and open source,” Keller said. How does the chip veteran foresee the future of AI chips? Here are the highlights from the Q&A session with Keller.
As the head of an AI chip startup, do you frequently use AI chatbots like ChatGPT?
“Actually, I rarely interact with AI. There’s no special reason for this. I spent years designing chips for video games in the past, but I didn’t play video games. Instead, I read books.”
With the rapid advancement of AI, there’s a heated debate between “Doomers,” who advocate for slowing down AI development, and “Boomers,” who support actively utilizing AI. Which side are you on?
“I’m definitely not a Doomer (laughs). I’m closer to a Boomer. AI is an incredibly fascinating and disruptive technology. However, humans have previously managed to handle disruptive technologies quite well. We already live in a world with intelligence disparities, wealth gaps, and differences in corporate sizes. AI can exacerbate these differences but also create new opportunities for humanity. Let’s wait and see how things unfold.”
You recently predicted that AI will replace all traditional software.
“Yes, I’m confident about that. Just as CD-ROMs disappeared and smartphones replaced computers, AI will rapidly replace traditional software. In the past, humans wrote thousands of lines of code for voice recognition, but now AI recognizes speech through neural networks. Traditional programs will be replaced by AI-generated ones, with humans reviewing them. Today, perfecting self-driving car technology takes millions of hours, but AI can reduce that time, just as my daughters learned to drive in 30 minutes without needing to drive millions of miles. The change has already begun.”
Many young engineers want to work in AI. What advice would you give them?
“AI is a very new field right now, offering a lot to learn. It’s a great time to dive in. However, engineers at large companies often focus on creating new versions of old products, which makes it hard to learn new things. This might be okay for 1-2 years, but eventually, you’ll need to seek new challenges.”
How should we educate the next generation in the face of rapid AI development?
“We should teach arts and fundamental sciences. High schools are teaching programming, and universities are teaching CAD, which is madness. We should focus on reading, writing, thinking, performing arts, music, and learning mathematics, physics, chemistry, biology, and history. Basics are always the best. I never studied computer science or architecture in college. Instead, I learned physics, electrical engineering, and 1850s electromagnetism theory, as well as calculus, astronomy, biology, and philosophy from the 1600s. I read ancient Greek philosophy and ‘The Art of War.’ I have dyslexia, so writing is difficult for me, but I loved math. The 18th and 19th-century mathematical concepts like Laplace transforms and Fourier transform were among the best things I learned. They helped me not only to think but also to find what I love.”
“Open source always wins”
Keller joined Tenstorrent in late 2020 after leaving his role as Senior Vice President at Intel. Initially joining Tenstorrent as its CTO, he was promoted to CEO in January last year. The company is currently preparing to launch Quaxar, the successor to Grayskull, which will be manufactured at a new Samsung foundry in Texas.
When is Quaxar expected to be released?
“We’re aiming for the end of this year. Our goal is to market these semiconductors to a broad range of customers, including chip developers, data centers, software developers, and general manufacturing companies. AI semiconductors will be integrated into more products, meaning our potential customer base is very diverse.”
Why did you choose Samsung Electronics Foundry to produce Quaxar instead of TSMC?
“We have experience working with both Samsung and TSMC. Previously, TSMC manufactured our 6-nanometer (billionths of a meter) semiconductors, while Samsung will produce our next-generation 4-nanometer Quaxar. Our decision was based on a combination of each company’s technology level, readiness, IP availability, pricing, and design support. Samsung has been a great partner. I’ve been collaborating with Samsung since 1994. Of course, TSMC is also a trusted partner. Both companies have earned my confidence.”
Why did you recently open a Korean office?
“To better support our Korean customers. We currently have a small team in Korea, but we plan to hire more technical talent soon. There are many excellent engineers in Korea.”
Tenstorrent’s semiconductors are based on the RISC-V architecture. Do you think RISC-V can outperform the architectures of major companies?
“Absolutely. Currently, AMD and Intel monopolize their architectures and restrict access to other companies. Using AMD’s architecture also requires significant financial investment. In contrast, RISC-V is open source and accessible to everyone. Currently, around 20-30 companies are using RISC-V technology. As a result, we expect RISC-V to be featured in chips from Intel, AMD, and Nvidia, as well as in the chips used by companies like Google and Meta.”
Is this the power of open source?
“Yes, it is. Open source always wins. When many people collaborate, innovation flourishes. RISC-V will become the standard for the next generation of architectures.”
RISC-V was developed as an open-source ‘Instruction Set Architecture (ISA)’ by UC Berkeley in 2010. ISA refers to the set of commands that a computer’s CPU (Central Processing Unit) can understand and execute. Unlike ARM’s ISA, which requires a license purchase, RISC-V is available for free. This allows anyone to design and use CPUs freely.
The AI semiconductor market seems to be dominated by NVIDIA. What is Tenstorrent’s strategy?
“First, let’s consider the entire AI market. It is not entirely dominated by NVIDIA. The AI market is very large. Many aspects of AI are run on Intel and AMD computers through the cloud. Apple uses AI on iPhones and Macs.
While NVIDIA does have a competitive edge in the high-performance chip market, that is not Tenstorrent’s target market. Our products differentiate themselves with excellent cost-performance ratios. We do not directly compete with NVIDIA; instead, we create products for markets with different needs. We design processors that operate at lower power and have the advantage of easily expandable technology through open-source.
We also provide AI engines and CPU technology in a licensing model, allowing anyone to create their own products. While media focuses on NVIDIA due to its high market cap, there are plenty of opportunities in other markets as well.”
Global tech giants like Meta and OpenAI are producing their own AI chips. Isn’t this a threat to traditional semiconductor companies?
“It isn’t necessarily a threat. For instance, AMD and Intel have diverse business areas. Although I don’t know the exact ratio, approximately 30% of AMD and Intel’s business might be related to large corporate clients. They also have small businesses and consumers as customers who do not produce their own chips.
Moreover, traditional semiconductor companies have made significant investments in many advanced technologies, not just AI. They can find new markets or other specialized areas. I believe they can overcome the current market changes. The AI semiconductor market is still evolving, providing opportunities for both new and existing players.”
Do you have plans for an IPO?
“Absolutely. We plan to go public within the next few years.”
Over the past few decades, Keller has left an indelible mark on semiconductor history with numerous groundbreaking projects. Yet, he never stayed at one company for long. After launching a revolutionary product, he moves on to another company, earning him the nickname “problem solver” in the industry. What is his secret to consistently creating innovative products?
You’ve created remarkable products at many companies, including AMD, Apple, Intel, and Tesla. While you could have stayed in one place and enjoyed your success, why did you move around so much?
“I simply love completing products. My goal is to find new opportunities and challenges. I made several products at Apple, but after Steve Jobs passed away, I became interested in new opportunities. At AMD, I successfully developed the Zen processor and its design team. At Tesla, I completed the self-driving chip ‘HW3′ in 18 months. This led to the creation of ‘HW4′ and the autonomous supercomputer ‘Dojo.’”
Does this mean you might leave Tenstorrent after completing another product there?
“Not necessarily. We already have several designed products and are creating something truly amazing. I have many ideas. My mission is to make Tenstorrent highly successful. There’s a lot to do, and I’m having a lot of fun with it.”
Is your goal to grow Tenstorrent into a company like NVIDIA, Intel, or AMD?
“Absolutely not. Customers already use products from NVIDIA, Intel, and AMD. I want to create a different kind of company. Musk didn’t set out to make an electric car like Toyota’s. He followed where electric cars led him and ended up creating a new kind of car.”
What are the key factors to becoming a winner in the AI semiconductor market?
“There are several. Curiosity, problem-solving skills, passion, and a strong work ethic are crucial. Especially in AI chips, performance is essential to stand out among competitors. When buying computers, we typically consider performance, programmability, availability, business terms, licensing, and power consumption. Any company entering this market must excel in at least one of these areas.”
As AI becomes more advanced, will the traditional computer structure, the ‘Von Neumann architecture.’ composed of processing and storage devices, also change?
“The three major categories of computer technology have not changed for a long time: computation, memory, and input/output (I/O). Among these, computation is the most complex part. Modern computers have three types of computation: CPU, GPU (Graphics Processing Unit), and TPU (Tensor Processing Unit, a semiconductor dedicated to AI). However, the paradigm is shifting. In the past, 90% of computer computation was done by the CPU and 10% by the GPU. In the future, the CPU will handle 10%, the GPU 10%, and the TPU 80%.”
You have worked with tech industry magnates like Steve Jobs and Elon Musk. Based on your experience, what advice would you give to future tech leaders?
“Truly enjoy your work. When I was young, a friend’s wife joked that we must put something in our drinks because we worked all day and talked about work all night. We loved our work and were extremely passionate about it. If you don’t like what you’re doing or you don’t get along with your boss, quit that job and find a place where you can work with truly passionate people.”
Finally, what is your secret to consistently being innovative?
“There is no secret. I just focus on the tasks and problems that need to be solved. While many people work for promotions and money, that’s not the case for me. My interest lies solely in solving problems. Additionally, it’s important to be very open-minded. If you don’t know something, ask questions. Since I was young, I would ask questions if I didn’t understand something. Some people avoid asking questions to appear smart, but then they learn nothing. This is a common mistake among young people. Don’t try to look smart; try to learn.”
