Dnotitia demonstrates dedicated vector-search ASIC

Dnotitia demonstrates dedicated vector-search ASIC

Dnotitia has demonstrated dedicated silicon for large AI retrieval workloads. Its VDPU ASIC searched approximately 2.6 million video frames at SC26, with the company reporting more than 90% lower host CPU utilisation for the tested search.


IN Brief:

  • Dnotitia has demonstrated its application-specific VDPU silicon at Supercomputing 2026 in Chicago.
  • The processor searched approximately 2.6 million video frames, with at least 90% lower host CPU utilisation reported for the demonstrated workload.
  • Further server-scale evaluations will examine integration with existing vector databases, storage systems, and AI infrastructure.

Dnotitia has demonstrated its Vector Data Processing Unit (VDPU) application-specific integrated circuit performing live semantic searches across approximately 2.6 million video frames at Supercomputing 2026 in Chicago. The South Korean company reports that offloading the search workload to dedicated silicon reduced host CPU utilisation by at least 90% compared with CPU-based execution of the same task.

The demonstration marks a transition from earlier FPGA-based evaluations to operation on the company’s purpose-built ASIC. Dnotitia previously presented its vector-processing architecture at the AI Infra Summit in September, while its first-generation silicon had entered evaluation during 2026. The latest demonstration establishes a functioning implementation using the dedicated processor rather than programmable logic configured to reproduce the proposed architecture.

Vector search is used in AI systems to retrieve information according to numerical representations of its content. Documents, images, and other data can be converted into vectors, allowing search software to identify entries with similar characteristics without relying entirely on exact keyword matches.

This approach is used in semantic search and retrieval-augmented generation, where an AI application obtains relevant information from an external knowledge source before producing a response. As the volume of stored material increases, comparing query vectors with large collections of indexed data can place substantial demands on computing resources.

Dnotitia’s demonstration involved natural-language queries over a collection representing approximately 2.6 million video frames. The system searched the associated vector representations to identify relevant content, with the VDPU performing operations that would otherwise consume host processing resources.

The processor has been designed specifically for these retrieval operations. Rather than using a general-purpose CPU to manage the complete search, the architecture moves selected vector-processing work onto an accelerator card equipped with dedicated silicon and memory resources.

Keeping the vector index close to the processor can reduce repeated transfers of information between host memory and the search engine. The approach also allows the host CPU to continue handling application execution, request management, and other computing tasks while the accelerator processes retrieval operations.

Dnotitia reports host CPU utilisation reductions of at least 90% during the demonstrated search workload. That figure concerns the processing resources used by the tested operation and should not be interpreted as a 90% reduction in total server electricity consumption, operating costs, or application latency.

The company has also outlined a target of approximately ten times the vector-search performance of a dual-socket CPU server when using a system with multiple VDPU cards. This remains a development objective rather than an independently verified performance result established by the SC26 demonstration.

The distinction is important because the performance of a retrieval system depends on more than the speed of individual vector comparisons. Search algorithms, index construction, memory capacity, data movement, query concurrency, and the required accuracy of returned results can all affect the final application.

Different vector-search workloads also impose different demands. A system processing a largely static video index may behave differently from one where documents are added, updated, and removed continuously. The balance between building an index, maintaining it, and executing searches determines how effectively acceleration can be used.

Dnotitia develops its Seahorse vector database alongside the VDPU hardware. The company has also described work with widely used vector-search software, including Milvus, FAISS, and hnswlib, allowing the processor to be evaluated within established retrieval environments.

Compatibility with existing software is a practical requirement for introducing a specialist accelerator into production infrastructure. Database applications must be able to direct suitable operations to the hardware while maintaining the indexing and retrieval behaviour expected by the wider application.

The company envisages computing systems in which CPUs handle general application functions, GPUs perform model inference, and the VDPU executes vector retrieval. Such an arrangement separates workloads according to their processing characteristics, potentially reducing competition for CPU and GPU resources within an AI system.

Retrieval-augmented generation is one prospective application. An AI agent may repeatedly search external data while completing a task, retrieving documents or other information that the language model cannot obtain from its existing context. The resulting search operations can become a separate computing workload within the infrastructure.

The benefits of dedicated hardware will depend on how effectively those requests are transferred to the accelerator, the size and structure of the indexed data, and the work required elsewhere in the system. Faster retrieval processing does not automatically improve every part of an AI application’s response time.

Dnotitia is continuing server-scale evaluations with partners across computing, memory, storage, and vector database technology. These assessments will establish how the ASIC performs under different system configurations and operational workloads.

The company has not confirmed general production shipment volumes, a broad commercial release date, or independently certified performance results for the completed accelerator. The SC26 demonstration establishes live ASIC operation and a company-reported reduction in host CPU utilisation, with wider performance and deployment characteristics still subject to evaluation.


Stories for you


  • Dnotitia demonstrates dedicated vector-search ASIC

    Dnotitia demonstrates dedicated vector-search ASIC

    Dnotitia has demonstrated dedicated silicon for large AI retrieval workloads. Its VDPU ASIC searched approximately 2.6 million video frames at SC26, with the company reporting more than 90% lower host CPU utilisation for the tested search.


  • Bel Fuse extends 10GbE copper transceiver reach

    Bel Fuse extends 10GbE copper transceiver reach

    Bel Fuse has introduced a longer-reach copper Ethernet transceiver module. Its new SFP+ device supports 10GBASE-T connections up to 100 metres, with the manufacturer reporting 35% lower power consumption than the previous generation.