Artificial intelligence infrastructure is usually discussed in terms of model compute.
Large Language Models need GPUs. Edge systems need accelerators. Training requires large clusters.
Retrieval introduces another requirement: finding the right information quickly enough to feed those models.
During our visit to Dnotitia in Seoul, CTO Se-Hyun Yang explained why the company believes vector search deserves dedicated hardware.
How vector search fits into AI
Modern AI systems frequently represent information as vectors.
An embedding model converts text, images, video or audio into numerical representations containing semantic information. Similar concepts occupy nearby regions in that vector space.
When an application receives a query, the system can convert that query into another vector and search for the closest matches.
This forms the retrieval layer used in applications including semantic search, recommendation systems, Retrieval-Augmented Generation and AI agents.
The challenge appears when the index becomes large.
Why CPUs and GPUs can struggle
Vector search creates a memory-heavy workload with irregular access patterns.
A CPU may spend significant time waiting for data to arrive from memory. GPUs provide high memory bandwidth, but GPU memory is valuable capacity that may already be required for embeddings or model inference.
Scaling retrieval by adding more general-purpose processors can therefore increase system cost and resource usage.
Dnotitia developed the VDPU around this specific workload.
The accelerator keeps the vector index in memory on the PCIe card and performs the search close to the data.
Queries move across PCIe, the card searches the index, and only the relevant matches return to the host.
Separating retrieval from inference
VDPU does not replace the GPU running a language model.
The architecture separates responsibilities.
GPUs or CPUs continue to generate embeddings and run model inference. VDPU handles vector index construction and retrieval.
This allows model compute and retrieval to scale independently.
That separation becomes particularly relevant for AI agents, where a single user request may trigger several searches before the model produces a final response.
Seahorse and the software layer
Dnotitia also develops Seahorse, its vector database platform.
Seahorse is designed around semantic retrieval and Retrieval-Augmented Generation workflows, including multimodal data and cloud deployment.
The wider software platform gives Dnotitia a route from the vector database layer down to its own acceleration hardware.
For engineers already using established software, the VDPU currently supports Milvus, FAISS and hnswlib.
Beyond the data centre
The interview also covered Mnemos, Dnotitia’s personal Large Language Model device.
Mnemos is aimed at running language models locally rather than relying entirely on cloud infrastructure. Dnotitia positions it for environments including offices, factories, robotics, government and medical systems.
The same broader architecture applies: local models still need access to relevant data, and vector retrieval provides the memory and search layer around those models.
Building language models around Korean
Dnotitia also develops the DNA family of language models.
The company initially focused heavily on Korean-language performance while maintaining English support.
Its published DNA 1.0 model uses an eight-billion-parameter Llama-based architecture and is available openly for further development.
This gives Dnotitia a stack spanning models, data infrastructure and dedicated retrieval silicon.
Evaluating VDPU
For engineers interested in evaluating the hardware, Dnotitia currently runs proof-of-concept engagements using customer data.
The process involves providing an index or representative sample and a query log. Dnotitia then measures throughput, latency and recall and compares the resulting configuration against the existing infrastructure.
The commercial ASIC is still moving through evaluation. First-generation VDPU samples returned from fabrication in July 2026 and are undergoing chip-level testing.
For teams building large RAG systems, recommendation engines or agentic AI platforms, the important point is that retrieval now has its own scaling problem.
Dedicated vector hardware is one approach to solving it.
Comments are closed.
Comments
No comments yet