Back in 2017, the idea of launching a startup to make artificial intelligence (AI) semiconductors prompted many blank stares. It was a fight that couldn’t be fought against global IT heavyweights.
But that’s exactly what Baek Jun-ho and Kim Han-joon, both researchers at Samsung Electronics, did. The two, who designed semiconductors at AMD and Samsung Electronics, were confident that AI chips would be the next big thing.
The following year, 20 other professionals with extensive backgrounds in semiconductors from Samsung Electronics, Qualcomm, AMD and Korea Advanced Institute of Science and Technology left their stable jobs to join the adventure. And this was the beginning of Furiosa AI, the country’s only fabless semiconductor startup that makes neural processing units (NPU).
WORTH A SHOT EVEN FOR STARTUP
Furiosa AI is tapping into the AI semiconductor market that powers servers for data centers and autonomous vehicles. This market is led by Nvidia, Intel, Google and other global giants. The Korean semiconductor industry is heavily concentrated on memory chips. It takes up only 3% of the non-memory semiconductor market, which requires logic, calculation and control for complicated data processing.
Even among non-memory semiconductors, AI chip is considered to be highly challenging. That’s because AI chips play the role of a brain inside the entire AI infrastructure. There are 10 billon to 20 billion transistors that move independently while simultaneously working together to power the AI algorithms.
The majority of AI chips that are currently on the market remain in the “training” stage, where they have completed a certain level of algorithm based on learning a massive amount of data. The next stage is “inference.” This is where a single chip is equipped with the ability to make a knowledgeable decision based on accrued information.
Think how autonomous vehicle works. In an unexpected situation, the vehicle decides whether to slow down or make a complete stop. This chip is currently in the research stage. Israeli startup Habana Labs put out a prototype and that’s about it. Inference chip is what Furiosa AI is targeting. It plans to release a prototype next year and compete against global players.
Many advised to go for the Internet of Things (IoT), which is comparatively less competitive. However, Baek stressed: “AI chips is actually the market that is worth a shot for startups.” He explained that the industry fits well into the nature of a startup environment where a small number of people make quick decisions and prompt executions. He added: “The central processing unit (CPU) market led by Intel may be impossible to break into, but data center and self-driving chips are a whole new domain. We’re at the same starting point as Nvidia and Google.”
NATIONAL LEADER IN AI CHIPS
A good balance between utility and universality is a necessity in order to succeed with AI chips. If they are designed to work only within a particular application, the price may drop but its universality also drops. At the same time, if the chips are designed to work with a wide range of applications, the price inevitably goes up. Furiosa AI’s goal is to lower the price of chips, while enhancing its universality.
This startup will participate in the global AI benchmark MLPerf next month. The event is considered the Olympics of technology as AI chip makers will be competing its performance against each other. The evaluation will be based on how fast and accurate the chips classify images and detect targets, among other categories. In Korea, Samsung and Furiosa AI are the only two that were invited. Among the 15 participating companies, Furiosa AI is one of only three startups.
Furiosa AI’s company name was inspired by the movie “Mad Max.” She is a figure who overcomes all kinds of adversity and obstacles and still manages to open a new era. “Developing an AI chip in Korea may seem like a David and Goliath battle, but our bold attempt itself will become a meaningful asset to the Korean semiconductor ecosystem,” said Baek.
By Suyoung Jo