A Korean Startup Put Vector Search on Its Own Chip: Dnotitia's Retrieval ASIC Runs Live for the First Time, Searching 2.6 Million Video Frames by Plain-Language Query
Most AI chip announcements compete on compute: how many FLOPS, how much HBM, how many accelerators per rack. Dnotitia, a Seoul-based company, is going after a different job. Its silicon does not do the math — it does the looking.
On November 19, at the SC26 supercomputing conference in Chicago, Dnotitia publicly ran the ASIC version of its VDPU (Vector Data Processing Unit) for the first time. A visitor typed a question in plain language, and the chip searched roughly 2.6 million video frames in real time to surface the clips closest in meaning.
It was the first public demonstration of the chip in application-specific integrated circuit form. In September, at AI Infra Summit 2026, the company showed an FPGA version — reconfigurable silicon used for validation. Moving to an ASIC means the logic is now fixed in custom silicon, which is a step closer to a shippable product.
This article is compiled from multiple media reports: the official press release distributed via PR Newswire, HPCwire, Korea's DigitalToday and Seoul Economic Daily.
TL;DR
- The problem is retrieval, not compute. AI agents look things up constantly — reference material, a user's past activity — and that semantic search has been running on CPUs.
- What was demonstrated: a single VDPU ASIC card handling plain-language searches across about 2.6 million video frames in real time.
- The headline number from the company: offloading the same search workload from CPU to VDPU cuts host CPU utilization by 90% or more.
- The stated target: a single server fitted with multiple VDPU cards delivering roughly 10 times the search performance of a dual-socket CPU server. That is a goal, not a measured result.
- The division of labor: CPUs run services, GPUs run model computation, VDPUs handle vector search.
- What is still missing: price, power draw, process node, throughput figures, production timing and customer names. None have been published.
Why "looking things up" turned into a bottleneck
Start with what vector search actually is. To answer a specific question, a model often has to pull relevant material out of a corpus first. Keyword matching is brittle — the same idea can be phrased in completely different words. Vector search converts documents, images and video frames into strings of numbers, arranged so that similar meanings sit close together. Finding relevant material becomes finding nearest neighbors.
That workload is CPU-heavy. In an October 7 talk in Seoul, Dnotitia CEO MK Chung put a number on it: 60% to 70% of CPU load in data centers goes to locating data. He described a second pressure point as well — the KV cache models use as short-term memory keeps growing, no longer fits inside expensive high-bandwidth memory, and gets shuffled in and out, adding latency.
Agents multiply the problem. Humans used to do the searching; now a single task has an AI searching dozens or hundreds of times. The more lookups, the more CPU gets consumed by retrieval, leaving less for the services that CPU was bought to run.
What the demonstration actually ran
The setup was concrete: a library of about 2.6 million video frames, a user question in natural language, and a chip that compares frame vectors against the question vector to return the closest matches in real time.
The work ran on a VDPU card built around an ASIC. The company reports two outcomes: better search performance than a CPU-only approach, and host CPU utilization down by 90% or more. Those freed cores can go back to running agents or other services — which is the real pitch. The point is not a modest speedup on search; it is handing an entire block of CPU capacity back to the business.
Dnotitia also describes a three-way split: CPUs execute services, GPUs handle model computation, and VDPUs take vector search. The target is roughly 10 times the search performance of a dual-socket CPU server from a single machine fitted with multiple VDPU cards.
How this differs from September
Same demonstration family, two generations of hardware, and the difference matters:
- September (AI Infra Summit 2026): an FPGA-based VDPU, showing that the approach works and at what performance level. FPGAs are fast to modify but too costly and too power-hungry for volume.
- November (SC26): an ASIC-based VDPU, showing the design still works once frozen into custom silicon. This is the necessary step between validation and product.
Korean media described the SC26 showing as the first public disclosure of ASIC-based VDPU performance, matching the official release.
The software is not an afterthought
A chip that only shuttles data around is not usable on its own. Dnotitia builds its own vector database, Seahorse, along with agent software, and says it fed requirements from real search workloads into the chip's design. The company is also evaluating the VDPU for integration into third-party server and storage systems — in other words, it intends to sell accelerator cards and software rather than complete machines.
For background: the company raised a $61.2 million Series A in August, is headquartered in Seoul and runs a North America team out of San Jose. Its Seahorse storage also took first place in MLPerf's vector database benchmark in September. That is context, not the news.
Timeline: a demo is not an order form
- August 2026: $61.2 million Series A closed (background).
- September 18, 2026: FPGA-based VDPU and a four-card server shown at AI Infra Summit (earlier milestone).
- October 7, 2026: CEO speaks at AI Festa 26, citing the 60–70% CPU-share figure and claiming search up to 15 times faster than CPU.
- November 17–19, 2026: SC26 booth 5613, ASIC-based VDPU running live (this announcement).
- Production date, pricing, customers: not disclosed.
How to use this: if you are evaluating the approach
The first step is to measure your own retrieval share. Ask platform or ops teams for one number: within inference services, how much CPU time goes to model pre- and post-processing versus vector lookup and reranking. If retrieval is under about 20%, a dedicated card will likely cost more in operational complexity than it saves.
Second, decide whether your problem is slow retrieval or retrieval eating your CPUs. These are different failures. The first shows up in P99 latency; the second in CPU utilization and queue depth. Hardware like this mainly addresses the second.
Third, pin down the software path. An accelerator card has to plug into the vector store you already run — Milvus, Qdrant, pgvector and the like — or you are effectively replacing your data layer. Public information so far covers Dnotitia's own Seahorse and a general statement about evaluating third-party server and storage integration; there is no published list of supported interfaces.
Who it is for
- Teams running their own RAG or agent platforms at high query volume, especially multi-step agents that retrieve repeatedly within a single task.
- Video and image libraries. The demo itself was frame retrieval; surveillance review, asset search and multimodal knowledge bases are the same shape of workload.
- Data centers where retrieval already saturates CPUs and adding another row of CPU servers is the alternative being weighed.
- Server and storage vendors. The company says explicitly that it is looking for partners to validate integration.
When not to bother
- Low query volume. Tens of thousands of lookups a day fit comfortably on CPUs; a dedicated card just adds a failure point.
- Projects that need to buy now. With no production date, price or reference customer, no procurement process will clear.
- Teams locked into a specific vector database ecosystem where the interface does not exist yet — migration cost will dwarf hardware savings.
- Anyone reading this as a GPU substitute. The chip does not run models and cannot replace a GPU for inference.
Common mistakes and pitfalls
- "90% less CPU" is not "10 times faster." One figure describes CPU capacity freed; the other is a performance claim, and the company has not published a single unified speedup number.
- The 10x figure is server-level, and specifically a multi-card machine against a dual-socket CPU server — not one card against one core.
- The 15x figure comes from an October talk, a verbal claim about search acceleration, and is not the same measurement as the 10x server target. Do not blend them.
- A demo is not production. An ASIC running proves functional correctness; yield, packaging, drivers and certification all still lie ahead.
- There is no third-party benchmark. Every number is vendor-supplied, with no test methodology, dataset size, CPU model or concurrency level disclosed.
Cost and timing
No price or delivery date has been published, so here is a framework rather than invented figures:
- Cost: a card has to cost less than the CPU capacity it replaces to make sense. With current disclosure, per-card pricing cannot be estimated.
- Integration: connecting to a homegrown vector store is a multi-month engineering effort in conservative terms; connecting to an existing one depends on whether an adapter exists.
- Availability: this was a live demonstration. The company says it will run proofs of concept with global server and storage companies to assess performance and reliability. PoC to volume production usually takes one or more additional hardware spins.
Next step
The genuinely new thing here is not a benchmark number. It is that retrieval is being treated as a workload that deserves its own silicon, and that the design has moved from FPGA to a working ASIC. It is far from purchasable, but for agent platforms with heavy retrieval load, it belongs on the technology radar — revisit when pricing and production timing appear.
Sources
- Dnotitia official release via PR Newswire (Nov 19, 2026): https://www.prnewswire.com/news-releases/dnotitia-presents-first-live-demonstration-of-its-vdpu-vector-search-asic-at-sc26-302945556.html
- HPCwire (Nov 19, 2026): https://www.hpcwire.com/off-the-wire/dnotitia-presents-first-live-demonstration-of-its-vdpu-vector-search-asic-at-sc26/
- DigitalToday English edition (Nov 19, 2026): https://www.digitaltoday.co.kr/en/view/127786/dinotisia-unveils-first-live-run-of-asic-based-vdpu-cuts-host-cpu-use-by-over-90
- Seoul Economic Daily English edition (Nov 19, 2026): https://en.sedaily.com/business/2026/11/20/deepauto-rival-dnotitia-demos-ai-search-chip-at-sc26
- ZDNet Korea (Oct 7, 2026, CEO remarks): https://zdnet.co.kr/view/?no=20261007172520