A Guth Labs publication ↗
Back to News

Dnotitia demonstrates a chip built for AI vector search

A sage processor linked by gold traces to three cream video-frame tiles on a dark green board.
AI-generated editorial illustration

Dnotitia announced on November 19 that it had given the first public live demonstration of its Vector Data Processing Unit chip at Supercomputing 2026 in Chicago. The chip searched about 2.6 million video frames in real time in response to natural-language queries, the company said.

The accelerator compares numerical representations of video frames with a query to find relevant scenes without needing an exact keyword match. Dnotitia had shown performance on a reprogrammable FPGA in September. The Chicago demo used a card built around its dedicated chip instead.

Dnotitia said the accelerator cut host CPU utilization by 90% or more compared with running the same search workload on the CPU alone. That is a company-reported result for the demo, not evidence of the same savings across workloads. The company is evaluating integration with third-party server and storage systems and did not announce a sale date or price.

A dedicated search card could leave more CPU capacity for application work when AI systems need to retrieve data repeatedly.

Sources and citations

Revision history

Published .