Dnoticia Demonstrates Performance of 4-Card VDPU Server in the U.S.
IT DAILY ·
✦ AI Summary
Dnoticia said it unveiled a server-scale expansion architecture for its vector search processor, the VDPU (Vector Data Processing Unit), at the AI Infra Summit 2026 held on the 18th at the Santa Clara Convention Center in California.
At the event, it demonstrated a server equipped with 4 VDPU cards, and CTO Yang Se-hyun presented the VDPU architecture and performance evaluation results under the title "Rethinking AI Infrastructure with Dedicated Vector Silicon."
In an FPGA-based evaluation, the server with 4 VDPU cards delivered vector search throughput up to 5.77 times higher based on a 4096-dimensional multimodal workload, while host CPU usage fell by 92% and memory use dropped by 73%.
Dnoticia announced on the 18th that it had unveiled a server-scale expansion architecture for its vector search processor, the VDPU (Vector Data Processing Unit), at the AI Infra Summit 2026 held at the Santa Clara Convention Center in California.
Following its unveiling of VDPU chips and accelerator cards at the previous FMS (Future of Memory and Storage) 2026, the company expanded the scope this time to applying them in a real server environment. On site, it demonstrated a server configuration equipped with 4 VDPU cards.
The presentation was delivered by Yang Se-hyun, CTO of Dnoticia. Under the title "Rethinking AI Infrastructure with Dedicated Vector Silicon," he introduced the VDPU architecture and performance evaluation results.
The first-generation VDPU ASIC is currently undergoing characterization to measure and verify chip operating characteristics. Dnoticia plans to begin ASIC-based evaluations from the fourth quarter of this year.
In FPGA-based verification, a server equipped with 4 FPGA-based VDPU cards was compared with a dual-socket CPU-only server running the same software stack. Based on a 4096-dimensional multimodal workload, the FPGA-based evaluation showed vector search throughput of up to 5.77 times higher. Host CPU usage for index building fell by 92%, and memory used for index building decreased by 73%.
During this process, search recall was equal to or higher than in the CPU-only environment. The throughput gain also showed that search quality was not compromised. The target for ASIC-based VDPU was then presented as up to 10 times the vector search performance of a CPU server.
Dnoticia explained that generation is handled by the GPU, while search is handled by the VDPU. The GPU's role is to focus on model execution and answer generation. The VDPU's role is to support external information retrieval and answer generation, and the deployment structure is responsible for vector search in retrieval-augmented generation (RAG) and AI agents.
Yang Se-hyun, CTO of Dnoticia, said that with the spread of agentic AI, the importance of data retrieval performance beyond model computation is growing. He added that the VDPU is dedicated to vector search, and that it was designed so the CPU and GPU can focus on other tasks.
Yang said that the company unveiled the VDPU server configuration at this event and plans to expand evaluation and deployment based on actual customer workloads going forward. Separately, Dnoticia's AI storage system Seahorse Storage ranked No. 1 in throughput in the MLPerf Storage v3.0 vector DB evaluation.
Source: IT DAILY · Kwon Young-seok
Original: https://www.itdaily.kr/news/articleView.html?idxno=241713
References
This article was produced with the help of an automated content generation algorithm.
Source: IT DAILY
View originalThis article was summarized and organized by BizCrush based on the original article from IT DAILY. For exact quotations and full details, please refer to the original article.