Dnotitia has received the first ASIC samples of its Vector Data Processing Unit and begun chip-level characterization, with evaluations of the custom silicon planned for the fourth quarter. At Santa Clara’s AI Infra Summit, it presented a four-card server for retrieval workloads.
The published performance results come from an FPGA evaluation platform, rather than the newly fabricated ASIC. Dnotitia says that platform delivered up to 5.77 times the vector-search throughput of a dual-socket CPU-only server running the same software stack, while maintaining equal or better recall.
For a 4,096-dimensional multimodal workload, the company reports reductions of 92 percent in host CPU use and 73 percent in host memory use during index building. Those measurements concern the FPGA configuration. Dnotitia’s target of up to ten times CPU-server vector-search performance for an ASIC-based server remains a development goal.
The FPGA platform has been validated with FAISS, Milvus and hnswlib, including brute-force nearest-neighbor search and several index structures. The company plans broader library and database support for the ASIC. It is discussing evaluations and integration with server, storage, memory and semiconductor companies, vector database providers and AI framework developers.
Dnotitia develops VDPU alongside its Seahorse vector database and storage software. The processor is designed for vector search and graph traversal, moving vector operations closer to stored data to reduce work on the host CPU and unnecessary data transfers. The company describes the hardware and software as a jointly designed system.
The server presentation follows Dnotitia’s public display of its VDPU chip and accelerator card at the Future of Memory and Storage event in August. Seahorse AI Storage received that event’s AI Application Award. The company used the display to seek storage-vendor partnerships for integrations and enterprise proof-of-concept projects.
Seahorse combines document processing, knowledge management, vector retrieval and AI-service integration. Dnotitia says it analyzes layouts, tables and charts, and combines meaning-based search with methods for finding exact terms, figures and clauses. Retrieved material can be linked to its source so users can check the evidence behind an AI answer.
The software supports cloud, on-premises and air-gapped deployment. Dnotitia is headquartered in Seoul and has a US office in San Jose, California. Its hardware program targets retrieval-augmented generation, in which an AI application obtains outside information to use in generating an answer.
Other AI data-path work includes Google’s TurboQuant research on compressing inference memory and Tower and NewPhotonics’ laser-integrated optical-chip shipments. VDPU targets vector retrieval, a separate workload from model execution and optical networking.
ASIC-based evaluations are scheduled for the fourth quarter of 2026.
Sources: Design & Reuse, Design & Reuse
–
By the Control Plane Editorial Team