Sapeon Adds AI to Edge Servers
After spinning off from SK Telecom, Sapeon has become the first Korean company to deliver an AI chip. The X220 provides an efficient accelerator for both video analysis and language processing.
New AI vendor Sapeon is a small startup wrapped in a $14 billion company. The 70-person team recently spun out of SK Telecom, the largest cellular-service provider in South Korea with 30 million subscribers. Sapeon markets a family of AI accelerators that the company originally developed for internal use, but now hopes to sell to a broader customer base. The initial X220 chip delivers 106 trillion operations per second (TOPS) at 65W, making it an efficient device for edge and data-center servers.
The X220 implements a custom AI-inference architecture comprised of a single large matrix engine for tensor operations as well as 16 vector engines that handle activation functions and similar operations. Despite the use of outdated 28nm manufacturing, it delivers strong performance per watt on several neural networks. Sapeon has been producing the X220 for internal use since 4Q20 and expects to ship the 7nm X330 next year.
Soojung Ryu is CEO of Sapeon; she served 15 years in Samsung’s R&D group before becoming VP of SK Telecom’s AI accelerator group. MK Chung heads the engineering team. The company is nominally headquartered in Silicon Valley, but Ryu and most other employees are in Seoul. SK Group owns 100% of Sapeon and provided 50 billion won ($37 million) in initial funding. The telco has used the X220 for some of its cloud services, including the Nugu voice assistant and T view video-surveillance system. NHN, a Korean broadband provider, also uses the X220.
Sapeon offers both single-chip and dual-chip cards based on the X220. The former is well suited to edge servers, meaning systems that are deployed at the edge of the network or even on customer premises. At 135W, the latter card targets cloud data centers. Initial use cases include real-time video upscaling, speech recognition, robotics, and video surveillance.