[DigitalToday reporter Daegeon Seok] Dinotisia unveiled for the first time a live run of its AI data search chip, the VDPU (Vector Data Processing Unit). The company said on Nov. 19 it carried out natural-language searches across about 2.6 million video frames at Supercomputing 2026 (SC26) in Chicago. The demonstration used a VDPU card equipped with an application-specific integrated circuit (ASIC) designed for a specific purpose.
Dinotisia said this was the first time it has disclosed the search performance of an ASIC-based VDPU. When a user searches in natural language, the VDPU searches and compares vector data from video frames to find information with similar meaning. The company showcased VDPU search performance in September at the AI Infra Summit 2026 using an FPGA, a chip whose internal circuitry can be changed even after manufacturing.
Search performance improved versus handling the same task with CPUs alone, and host CPU usage fell by more than 90 percent. The company said remaining CPU resources can be used to run AI agents or process other services after handing off searches to the VDPU.
AI agents repeatedly search for necessary materials and a user's past usage history while performing tasks. Semantic search, which finds information that matches the meaning of a question beyond word matching, is used for this. Vector search that implements it converts semantic features of documents or images into numbers and compares them, expressing information with similar meaning as numerically closer values.
Dinotisia is refining its server environment with the goal of delivering about 10 times higher search performance than a dual-socket CPU server on a single server fitted with multiple VDPU cards. It splits roles, with the CPU running services, the GPU handling AI model computations and the VDPU taking on vector search. It is developing the vector database Seahorse and agent software together, reflecting the demands of the search process in the design, and it is also conducting verification for use with external servers and storage to expand its AI storage business.
Dinotisia CEO Mu-kyung Jung (정무경) said, "Through this demonstration, we verified the search performance and actual operation of the first VDPU ASIC," adding, "We will continue proof-of-concept (PoC) work with global server and storage companies to verify performance and stability."