
Dnotitia demonstrates ASIC-based VDPU for vector search at SC26
The AMW Read
The ASIC demonstration advances Dnotitia's earlier FPGA work, but its wider impact depends on external performance and stability tests.
Dnotitia demonstrates ASIC-based VDPU for vector search at SC26
Dnotitia demonstrated its ASIC-based Vector Data Processing Unit (VDPU) at Supercomputing 2026 in Chicago. The demonstration used natural-language requests to search approximately 2.6 million video frames for semantically similar information. The company said the chip reduced host CPU usage by more than 90% compared with CPU-only processing of the same search task. The article does not provide the test configuration or detailed latency results.
The demonstration places Dnotitia in the market for specialized AI infrastructure that handles vector retrieval. Its approach assigns vector search to a dedicated chip, while CPUs run services and GPUs handle AI model computation. Dnotitia is also developing the Seahorse vector database and incorporating search requirements from its database and agent software into the hardware design. This integration could make retrieval performance and system compatibility central to its competitive position, though the demonstration alone does not establish an advantage across workloads.
For builders and investors, the next evidence is performance in external server and storage systems. Dnotitia said it is pursuing proofs of concept with global server and storage companies to test performance and stability. It is also optimizing a server with multiple VDPU cards, targeting roughly 10 times the search performance of a conventional two-socket CPU server. That figure is a target, not a demonstrated result.