"NVIDIA and FuriosaAI are, so to speak, engaged in an 'F1 (Formula One) race.' To survive in F1, one must achieve extremely high speed and performance."

Baek Jun-ho, founder of 'FuriosaAI,' stated this during an interview with Dong-A Ilbo at the company's headquarters in Gangnam-gu, Seoul, on the 14th, pointing beyond the window to Dosan-daero. He explained that if a city designed by an architect for 100,000 residents suddenly had to accommodate 100 billion people, the roads would become congested, losing functionality.

Baek emphasized, "Similarly, the volume of computational tasks that AI semiconductors must process has increased a million-fold in seven years, and our philosophy is that we must be able to run the most powerful existing AI models."

Baek, a former engineer at the American fabless company AMD and Samsung Electronics, founded FuriosaAI in 2017. FuriosaAI is a domestic fabless company developing high-performance chips specialized for AI inference. In March of this year, it rejected an acquisition offer from the American big tech company 'Meta' worth $800 million (approximately KRW 1.2 trillion) and became a 'unicorn' (a startup valued at over KRW 1 trillion) just over four months later. Previously, President Lee Jae-myung visited the headquarters during this year's presidential election period, stating, "FuriosaAI is providing hope to the public in the AI field."

He views the next three years as a golden time that will determine the company's fate. Baek stated, "The AI semiconductor infrastructure has a monopolistic nature. Due to the high technical and business complexity, survival becomes success and sets the standard. How we leverage the possibilities in the nascent semiconductor or model domains within the next three years will determine the outcome of the AI war."

He also cited a statement by Kevin Weil, Chief Product Officer (CPO) of OpenAI, who said, "The AI models we are using now are the lowest-performing models we will encounter in our lifetime." Baek noted, "In AI models, many innovations occur even within a few weeks' cycle," and added, "AI semiconductors must surpass the speed of model innovation. Through our third-generation products, we aim to target the market for running ultra-large models with hundreds of trillions of parameters on a massive scale."



The following is a Q&A with CEO Baek.


"I believed the AI era would come and founded the company eight years ago. The AI era requires a neural processing unit (NPU) optimized for AI computation, not a multipurpose graphics processing unit (GPU). It was an ambitious goal, facing many skeptical views."


"Of course, I didn't quit my job with the intention of starting a fabless company. While working at Samsung Electronics, I tore my Achilles tendon playing soccer and had to stay in the hospital twice. Having a lot of time in the hospital, I studied AI seriously. I gathered materials, including AI-related lectures from Stanford University in the US.

Unlike in the past, neural networks and machine learning had become very powerful. Then, in April 2016, I left Samsung Electronics. At that time, many good papers were published at global computing conferences, and since AI-native NPUs were a new area, I thought the entry barrier was relatively low compared to GPUs, making it worth trying."

"No, I didn't. However, being in my late 30s and unmarried, there was no one around to oppose my startup. I founded the company with Kim Han-joon, CTO, who also left Samsung Electronics, and Professor Koo Hyung-il from Ajou University, but since we didn't have an office, we held brainstorming meetings in Ajou University's seminar room. Then we saw an announcement from Naver D2SF, which invests in startups, and submitted an application. We received an investment of KRW 1.3 billion from Naver D2SF and started working in Naver's incubating space at Gangnam Meritz Tower. We sustained ourselves for 2 years and 9 months with KRW 1.3 billion."

"Although it was a relatively small amount, fortunately, good people gathered. About 25 initial members were highly skilled and trusted us, waiting patiently. There were times when salaries were not paid. Without production costs, we spent a long time writing the script, so to speak, in film terms. Thanks to that, we could write a very detailed script during the design phase. We laid the foundation by carefully observing AI workloads. Then, in the second half of 2019, we secured an investment of KRW 8 billion from the Korea Development Bank and others, which we used for our first production."

"With an investment of KRW 8 billion, we produced the first-generation chip 'Warboy.' At that time, it was produced at Samsung Electronics' foundry, and it came out in 'one shot.' (Producing in one shot means FuriosaAI's design is highly complete.) Although it was entry-level, we decided to evaluate its performance by global standards and submitted it to MLPerf, the most authoritative global benchmark, where it achieved very meaningful results. This gave us strong internal confidence that we could continue on this path."


"Currently, the demand for high-performance, high-efficiency inference infrastructure is global. There is a broad consensus among many companies that existing GPUs are not a reasonable choice due to various economic, social, and environmental issues."

"If GPUs are gasoline cars, we are electric cars. The design paradigm and category are different. We aim to be the 'next generation' as the global top in the NPU market. We are not trying to make gasoline cars like NVIDIA to compete, but there is a misunderstanding. Electric cars have their unique challenges, and the paradigm of technological innovation is different. Just as Tesla succeeded with new innovations, even big tech companies with their chips are showing great interest in our products and technology."


"When using Renegade, LG's high-performance conditions are met, and performance per power is improved by 2.25 times compared to existing GPUs. This means that while solving the excessive power consumption problem, which is a chronic limitation of existing GPUs, we achieved the specifications needed for large-scale generative AI services. We are currently testing with five large global customers, in addition to LG.

In a situation where major technology-leading countries are focused on securing their sovereign AI for AI sovereignty, the strategic value is high in that it reduces GPU dependency and allows AI infrastructure to be built with domestic technology. Last week, we conducted customer meetings in New York and received positive feedback. Renegade's AI model 'Llama' 70B driving performance exceeds that of NVIDIA's product line, receiving favorable reviews from customers."

"Our goal is to become the global leader in the NPU field within 3-5 years in the next-generation computing market. We already feel high expectations in the global capital market for our potential as the 'next NVIDIA.'"

"To create the world's best team, players with strong fundamentals like Son Heung-min and Lee Kang-in are needed. I believe FuriosaAI emerged because the fundamentals of Korea's semiconductor workforce, which is number one in memory semiconductors, were well-established. I believe there is a path. I think excellent Korean fabless companies can emerge."

Popular News