> ## Content Index
> Fetch the complete content index at: https://bytevyte.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Dnotitia VDPU First Samples Return From Fab as Vector Silicon Enters Characterization
- URL: https://bytevyte.com/dnotitia-vdpu-first-samples-return-from-fab-as-vector-silicon-enters-characterization/
- Published: 2026-09-20T12:03:45.000Z
- Updated: 2026-09-20T12:03:45.000Z
- Description: Dnotitia VDPU first ASIC samples return from fab, entering characterization with Q4 2026 silicon evaluations due after FPGA tests showed 5.77x throughput.
- Author: Bytevyte Editorial
- Tags: ai-beats

**Dnotitia** has brought its **Dnotitia VDPU** back from the foundry, reporting that the first ASIC samples are in hand and chip-level characterization has begun. The South Korean AI data infrastructure and semiconductor company disclosed the milestone on September 18 at AI Infra Summit 2026, where it framed the chip as dedicated vector hardware intended to scale from appliance-class systems to rack-level deployments. Dnotitia owns both halves of that stack: the Seahorse vector database in software and the Vector Data Processing Unit in silicon.

First silicon is a checkpoint rather than a product. It confirms that a design survived tape-out and that fabricated die can be powered up and measured, the step that precedes customer sampling and, eventually, volume production. Dnotitia says silicon-based evaluations are planned for the fourth quarter of 2026.

The device is not a general-purpose accelerator. Dnotitia designed the Dnotitia VDPU to parallelize graph traversal and vector similarity calculations, the work that currently falls to CPU servers sitting between storage and the GPU. In retrieval-augmented generation, a single user query can trigger an embedding lookup and a similarity search across a vector index before the model produces any output.

## Why Retrieval Has Become the Constraint

The dominant theme at AI Infra Summit 2026 was power rather than chip supply. Efficiency, power delivery and the economics of scaling compute now determine how much AI capacity an operator can actually run, and the retrieval layer sits directly in that cost path.

Large language models have limited ability to hold context across long sessions, which pushes production systems toward retrieval pipelines that fetch relevant passages from an external store at query time. Each fetch is a vector search. The volume of those searches scales with users rather than with training runs, which makes retrieval one of the few AI workloads whose demand rises directly with adoption.

Vector search fits poorly on general-purpose cores. It is memory-bound, latency-sensitive and repeated millions of times per query batch, and it consumes server power in proportion to how badly the hardware matches the workload. That mismatch is the commercial argument for dedicated silicon: cut the power spent per query and the cost curve for the entire inference pipeline moves with it.

Dnotitia's stated goal is to speed the delivery of data into GPUs and NPUs during inference, which places the chip in a slot between storage and the accelerator rather than in competition with either. The company has also said its VDPU intellectual property can be integrated into system-on-chip designs, so the same engine could appear on a card or inside a larger processor.

## What the Dnotitia VDPU Claims to Deliver

Every performance figure Dnotitia has published so far comes from FPGA prototypes, rather than from the samples now in the lab. The company reported up to 5.77 times the vector-search throughput of a CPU server in that configuration, and projects that a hardened ASIC implementation could lift vector database search performance by as much as 14 times. Dnotitia describes the second number as conditional on the FPGA results carrying over to production silicon.

| Metric                                  | Status                   | Figure              |
| --------------------------------------- | ------------------------ | ------------------- |
| Vector-search throughput vs. CPU server | FPGA prototype           | Up to 5.77x         |
| Vector database search performance      | Projected for ASIC       | Up to 14x           |
| Silicon-based evaluation                | Planned                  | Q4 2026             |
| ASIC version commercial release         | Previously stated target | Second half of 2026 |

The distance between those two figures is where the risk sits. FPGA results are provisional by construction: an FPGA runs at lower clock speeds and cannot match the memory subsystem of custom silicon, but it also avoids the yield, thermal and timing-closure problems that surface only after fabrication. Dnotitia has confirmed first silicon; a working product is a later step.

## The Trade-Offs Still to Clear

The immediate tension is timing. Dnotitia previously said the ASIC version of the VDPU would be released in the second half of 2026\. With samples only now entering characterization and evaluations penciled in for the fourth quarter, the window for a commercial launch inside that stated half-year is narrow. A slip into 2027 would not be unusual for a first-generation chip, but it would collide with the funding and go-to-market clock the company has set.

Capital is the second variable. Dnotitia closed a KRW 90 billion Series A, about 63.4 million US dollars, to expand its AI storage business, and it has been preparing for an initial public offering. That money buys design iterations and foundry capacity, and it also raises the bar for demonstrable results.

Characterization is where problems become expensive. A defect found in the lab can force a respin of the mask set, and each respin costs months that a startup selling into hyperscale procurement cycles cannot easily recover. Dnotitia's fourth-quarter evaluation schedule is as much a statement about confidence in the design as it is a product milestone.

Competition arrives from three directions. CPU vendors keep widening their vector instruction sets; GPU makers already accelerate similarity search inside their own libraries; and the data processing units sold by large networking vendors target the same data-movement problem with broader workloads in view. Dnotitia's bet is narrower, and narrower bets are easier to test: if a chip does only vector work, it should do that work at a power and cost per query that general-purpose parts cannot match.

Scaling adds another requirement. Vector indexes grow with the corpus they serve, so an accelerator that fits in an appliance can serve a department, while a rack-scale deployment has to hold indexes far larger than an appliance can address. Dnotitia has said the ASIC version would bring higher memory capacity and better power efficiency for hyperscale environments.

## What to Watch

For operators running large retrieval workloads, the decision point is the fourth-quarter evaluation data, not the announcement. A dedicated vector part pays off only if the reduction in power per query exceeds the cost of adding another card type to a rack, plus the software work of routing queries to it. Dnotitia's argument is that vertical integration lowers that integration cost, because one vendor controls the database and the chip underneath it.

That dynamic explains why the market produced data processing units and smart network cards before dedicated vector engines. Each generation of AI workloads has eventually pulled its most expensive, most repetitive operation out of the CPU and into fixed-function hardware, and vector similarity search is a candidate for the same treatment if the per-query power savings are large enough to justify a new part in the bill of materials.

For the wider market, the milestone tests a specific thesis: that AI infrastructure spending keeps shifting from the training cluster toward the serving path, where retrieval, memory bandwidth and power per query set the ceiling. Dnotitia is early to that thesis with silicon rather than a slide. Its Series A, its IPO preparation and its roadmap now rest on what the characterization lab reports back.

## Why This Matters

If the Dnotitia VDPU delivers on projections derived from FPGA testing, it would move part of the AI infrastructure cost base away from general-purpose CPUs and toward a purpose-built category that barely existed a few years ago. If it falls short, the result is another data point that dedicated accelerators struggle to beat incremental gains from incumbents. Both outcomes arrive within roughly two quarters, when evaluation results and any commercial release date become public.

## Related Articles

- [NVIDIA Starts Mass Production of Rubin R100 GPUs and Vera CPUs for Next-Gen AI](https://www.bytevyte.com/nvidia-starts-mass-production-of-rubin-r100-gpus-and-vera-cpus-for-next-gen-ai/?ref=bytevyte.com)
- [Avalue Technology Launches New High-Density Edge Solutions for Industrial AI and Medical Imaging](https://bytevyte.com/avalue-technology-launches-new-high-density-edge-solutions-for-industrial-ai-and-medical-imaging/)
- [NVIDIA Ships First Vera CPUs to OpenAI and Anthropic for Agentic AI Workloads](https://bytevyte.com/nvidia-ships-first-vera-cpus-to-openai-and-anthropic-for-agentic-ai-workloads/)

✔Human Verified

---

*Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.*