VDPU Vector Data Processing Unit

Purpose-built silicon for AI agent retrieval

VDPU is a PCIe card that offloads vector index building and search from your CPUs and GPUs. It works with the vector database you already run.

Request a PoC Explore VDPU

Every agent request starts with a search.Then it reasons, and searches again.One question turns into many retrievals.

CPUs sit waiting on memory.GPUs are busy serving the models.

Retrieval is the new AI bottleneck.

VDPU holds your vector index in on-card memory and searches it there. Only the results go back to the host.

Near-memory search: Compute sits next to the index on the card, so there are no round trips to the host.
Many queries at once: The card schedules and serves concurrent agent queries in parallel.
Plugs into your stack: Works through the Milvus and FAISS you already use.

How it works

Search moves onto a card in servers you already run. As agents add retrievals, you add cards instead of racks of CPU servers.

Less hardware for the same load

Fewer servers to buy, power, cool and maintain for the same retrieval traffic.

CPUs and GPUs freed for their real work

Host cores serve the application, and GPU memory stays with the models.

Steady responses under agent load

Latency holds steady as concurrent queries climb, so agents stay responsive.

Independently tested by a KOLAS-accredited lab: recall above 99%, with single-query p99 latency of about 5 ms. See the results

AI agents and deep research
Multi-step agents retrieve again and again within one request. VDPU keeps that loop fast under load.
Enterprise knowledge search
RAG over documents, logs and metadata, on-premises or in a private cloud.
Multimodal similarity search
Image and text embeddings searched together, at high recall.
Recommendation
Random-access lookups that stall CPUs, served from on-card memory.

Bring your index and query log. We'll measure throughput, latency and recall on your data, on your server or ours.

Request a PoC How a PoC works