Dnotitia announced the first public live demonstration of its VDPU vector-search ASIC at SC26 in Chicago on November 19. It describes natural-language searches across approximately 2.6 million video frames and at least 90% lower host CPU utilization when search is offloaded. These are company measurements. Approximately tenfold search performance in a multi-card server remains a target, not an established result of this demonstration.

What changes from the earlier stage

The company’s September release discussed FPGA evaluations and work on initial ASIC samples. That older release supplies background: the current development is a demonstration using the dedicated chip. A result from one programmable configuration should not be transferred automatically to different hardware or a future server installation.

Faiss documentation provides complementary software context. Similarity search compares vector representations and uses indexes to work with large collections. It is not an independent VDPU test. It helps explain why data selection, indexing and the definition of an acceptable result belong in the evaluation of such systems.

Freeing a CPU and shortening a response are different measures

Our interpretation is to measure host utilization separately from time to a useful answer. Work can move successfully to an accelerator while another stage still constrains the complete response. A CPU percentage should therefore not automatically become the same percentage of savings for the whole service. Offloading is a specific architectural change, not a complete application benchmark.

A hypothetical pilot could use a collection with known target frames and predetermined queries. The first check would establish that relevant results are actually found. Only then would speed, concurrent requests and resource use be compared. This is our proposed evaluation method, not a test performed by SIExpo.

What must remain comparable

The comparison becomes more informative with the same data and acceptance criterion. A configuration returning less relevant results more quickly is not an unambiguous winner. Different ranking approaches should be recorded rather than hidden inside one performance figure. Those details determine what the reported improvement means for an application.

Updates also matter when an agent works with changing information. A fast answer from yesterday’s index may be insufficient. Our proposed evaluation would add new material and check when it becomes discoverable. That is a question for future measurements, not a claim that VDPU has a known update problem.

The announcement documents a concrete hardware stage in the specialization of AI infrastructure. SIExpo has not reproduced the demonstration or established independent superiority. Comparable results and reliability during sustained operation would be the next useful evidence for assessing the product’s practical value.