Dinotisia Ushers in an AI Infrastructure Paradigm Shift: “Generation on GPU, Search on VDPU”
TECHWORLD ·
✦ AI Summary
Dinotisia unveiled an architecture based on VDPU, a processor dedicated to vector search, at AI Infra Summit 2026, extending it to a real server environment.
A server equipped with four VDPU cards was displayed at the event, and CTO Yang Se-hyeon presented the VDPU architecture and performance evaluation results.
In the evaluation, vector search throughput improved by up to 5.77 times on a 4096-dimensional multimodal workload, host CPU usage fell by 92%, and memory usage dropped by 73%.
Dinotisia said on the 18th that it had announced its participation in AI Infra Summit 2026 and unveiled an architecture based on VDPU (Vector Data Processing Unit), a processor dedicated to vector search. The newly disclosed setup has been extended to a real server environment.
Dinotisia had previously unveiled the VDPU chip and accelerator card at FMS (Future of Memory and Storage) 2026, and expanded the scope of application to a real server environment at this event. AI Infra Summit 2026 was held from the 15th to the 17th at the Santa Clara Convention Center in California, the United States.
At the event, Dinotisia set up a demonstration using a server equipped with four VDPU cards. Through this, it presented an architecture that extends VDPU (Vector Data Processing Unit), a processor dedicated to vector search, to a real server environment.
Yang Se-hyeon, CTO of Dinotisia, gave a presentation under the theme "Rethinking AI Infrastructure with Dedicated Vector Silicon." In his talk, Yang introduced the VDPU architecture and performance evaluation results.
For the evaluation, FPGA-based testing was conducted, with a server equipped with four VDPU cards set as the comparison target. The comparison target was a dual-socket CPU-only server running the same software stack. Based on a 4096-dimensional multimodal workload, vector search throughput improved by up to 5.77 times, host CPU usage for index building fell by 92%, and memory usage dropped by 73%. Search recall was equivalent to or higher than that of the CPU-only environment.
The company said these results showed that higher throughput and search quality were secured at the same time. In other words, it confirmed that search performance and resource efficiency improved while search quality was maintained or increased.
Based on these results, Dinotisia is working on characterization for its first-generation VDPU ASIC, and plans to begin full-scale evaluation in the fourth quarter of this year. The target for ASIC-based VDPU is up to 10 times the vector search performance of CPU servers. Dinotisia defined the roles as generation on GPU and search on VDPU. GPU focuses on running large language models and generating answers, while VDPU handles external information retrieval. Areas of application for VDPU include search for AI answer generation in retrieval-augmented generation (RAG) and dedicated vector search for AI agents.
As agentic AI spreads, repeated data retrieval and verification within a single task flow are becoming more common, making data search performance, in addition to model computation performance, a key element of AI infrastructure. Accordingly, within AI infrastructure, data retrieval capability is taking on greater importance alongside compute capability.
Dinotisia verified integration between Seahorse, FAISS, Milvus, and hnswlib and VDPU in an FPGA environment. This confirmed interoperability between VDPU and major vector search software in response to the spread of agentic AI.
For the domestic market, Dinotisia is pushing ahead with combining Seahorse and VDPU. Its domestic market goal is to productize an AI data infrastructure for AI storage, search, and utilization of enterprise-owned data.
For the global market, the targets are server, storage, memory, and semiconductor companies, and the global market strategy is to expand VDPU evaluation and proof of concept (PoC). CTO Yang Se-hyeon said that as agentic AI spreads, the importance of data search performance is rising alongside model computation, and explained that VDPU supports each of CPU and GPU focusing on their own roles based on a dedicated vector search design.
The speaker added that after unveiling the VDPU server configuration at this event, the company plans to expand evaluations and deployment based on real customer workloads going forward.
Source: TECHWORLD · Park Gyu-chan
Original: https://www.epnc.co.kr/news/articleView.html?idxno=407117
References
This article was produced with the help of an automated content generation algorithm.
Source: TECHWORLD
View originalThis article was summarized and organized by BizCrush based on the original article from TECHWORLD. For exact quotations and full details, please refer to the original article.