Sapeon Chief Technology Officer Michael Shebanow poses for a photo with the company's latest X330 compact card at the SK Tech Summit 2023 in Seoul on Thursday. (Sapeon)
Sapeon, a fabless startup backed by South Korean telecom-to-chip conglomerate SK Group, has come up with a more powerful and efficient artificial intelligence chip to compete head-on with US chip giant Nvidia, which enjoys a near monopoly in the soaring market for data center graphics processors.
“We‘re focused on being very good at (developing) AI inference (neural processing units) and we’ll take a part of the expanding (AI chip market),” Sapeon Chief Technology Officer Michael Shebanow said in an interview with The Korea Herald at the company’s office in Pangyo, Gyeonggi Province, Friday.
US-based Sapeon unveiled its new chip for data centers, dubbed the X330, during SK’s annual tech summit held in Seoul last week.
According to the company, the chip is four times faster and consumes half the energy compared to its first chip, the X220, launched in 2020. Its foundry partner, Taiwan’s TSMC, plans to start chip production using 7-nanometer manufacturing technology next year.
Sapeon's CTO stressed that the new chip is two times faster and 1.3 times more energy efficient when compared to a key rival in the market, without elaborating further on the details of the rival chip.
Competition in the AI chip market is heating up, with bigger rivals Nvidia and Microsoft being the key players. But Sapeon said it has achieved a competitive edge in lowering the total cost of ownership for its clients by using SK hynix’s more advanced GDDR6 DRAM chips and large language model.
While Nvidia, for instance, has focused on advancing chips overall for decades, elevating AI’s inference performance has been the 2-year-old Sapeon’s strategic focus. The CTO added that the ultimate goal is to make Sapeon a high-performance maker of processors like its powerful rival.
Sapeon was established as a joint venture with 80 billion won ($61.7 million) in funding from three SK companies -- SK Telecom, SK hynix and SK Square -- in 2021. The company then spun off from SK Telecom last year. The telecom carrier is still the largest shareholder with a 62.5 percent stake.
Currently, the firm has secured several Korean firms, including SK Telecom, NHN Cloud, Hana TI and broadcaster MBC as its key clients.
Shebanow, a renowned AI chip designer with over 35 years of robust experience, leads some 100 developers working at the US headquarters in Silicon Valley, California and the Korean office together as “one team.” Before joining Sapeon in June this year, Shebanow worked at several Big Tech companies, including Nvidia, Samsung Electronics and AMD.
“It's a privilege to work with SK Telecom. We’re a small company and we can’t go up against the world but SK Telecom can. Building an ecosystem is more important than what we do,” he said. “SK Telecom is globally competitive and we get to be part of that ecosystem. So that‘s a big advantage.”
Sapeon also joined an SK Telecom-led AI alliance, called K-AI Alliance, which was launched at this year’s MWC trade show in Barcelona, Spain. As part of the alliance, it seeks various ways to help companies offer more power-efficient AI inference services.
Market tracker Omdia predicts that the AI data center server market will grow at an annual rate of 31 percent between 2022 and 2027. The market for AI interference is expected to grow at a faster rate of 32 percent during the same period.
Following the X330, the company plans to launch an automotive AI chip, called the X430, at the end of 2025. It will likely be the company’s first chip to adopt HBM3 or high bandwidth memory 3, a next-generation, more advanced type of memory.
Shebanow also hinted at a new round of funding. Discussions with the firm's potential investors are underway for its Series B funding with the closing target planned for sometime next year. The firm raised 60 billion won in Series A funding in July 2022, which valued the company at 500 billion won.
By Jie Ye-eun (firstname.lastname@example.org)