explainx.ai350K+ learn from us
TrendingAI News TodayExplainx for Business
Pricing
explainx.ai

Upskill in AI — 16 free pathways, live workshops & bootcamps, and 50+ courses from practitioners. Plus the skills, tools, and MCP servers to practice on.

follow us

follow on google

Add explainx.ai as a preferred source

corporate training

support@explainx.ai

get started

Find your pathTake Free Evaluation

community

Join the community

learn

mind: share how you thinkpathways — start freeworkshopsbootcampscoursescompare Explainxcertificationsmock testsexplainx universitycorporate traininglearn skills & mcp

discover

skillsmcp serversexplainx mcpestacktoolsmdx readeragentsllmsdesignsdictionarypeopleagi trackerfelony benchranks

company

aboutvisionmissionteaminstructorsteach on explainxpartnershipscommunityhackathonscareers

content

daily AI newsstate of AI — live resultsblogreleasespromptsgeneratorsresource libraryfor LLMsexplainx.ai kids

solutions

all solutionsdeveloper upskillingmarketing upskillingproduct manager upskillingleadership upskilling

newsletter · weekly

Get AI news, tools, and insights in your inbox.

supportcontactprivacytermsdata rightshow we create contentsubmission guidelines

© 2026 AISOLO Technologies Pvt Ltd

explainx.ai

On this page

  • TL;DR: what NVIDIA announced
  • Why does an instrument need GPU-native acquisition?
  • What is the Berkeley Lab EELS camera?
  • What does hardware packet reordering change?
  • How much faster is the ibverbs engine?
  • What is direct polling and why 2-4 microseconds?
  • What this means for builders and instrument makers
  • Questions people will ask
  • Limits to keep in mind
  • How to try it
← Back to blog

explainx / blog

NVIDIA DAQIRI: Real-Time AI for Instruments, From a 500 Gbps Electron Microscope Camera

NVIDIA, AI for Science, Edge AI, Holoscan, Data Acquisition

Part of AI Chips and Infrastructure

NVIDIA DAQIRI moves instrument data into GPU memory during acquisition. A Berkeley Lab camera hits 500 Gbps; the new release adds NIC reordering and 2-4 µs polling.

Nov 19, 2026·8 min read·Yash Thakker
add explainx.ai
go deep
NVIDIA DAQIRI: Real-Time AI for Instruments, From a 500 Gbps Electron Microscope Camera

NVIDIA has published a technical post on DAQIRI, its library for sending high-bandwidth instrument data straight into GPU memory so software can analyze, and steer, an experiment while it is still running. The post, dated November 19, 2026, pairs a case study (a new electron microscopy camera at Lawrence Berkeley National Laboratory that streams more than 500 gigabits per second) with three new features: hardware packet reordering, an ibverbs engine and a direct polling mode. Every performance figure below comes from NVIDIA's own benchmarks.

TL;DR: what NVIDIA announced

table · 2 cols
ItemDetail (per NVIDIA)
What DAQIRI isA software-defined data acquisition library, part of the Holoscan platform, that moves sensor data into GPU workflows
Case studyHigh-dynamic-range electron energy-loss spectroscopy (EELS) camera from Berkeley Lab's Molecular Foundry
Camera data rateMore than 500 Gbps, over 500 times conventional EELS streams
Spatial resultSubmicron resolution from 5 micrometer sensor pixels
New: packet reorderingOffloaded to ConnectX-7 or newer NICs, payloads land in order in memory
New: ibverbs engineAbout twice the per-core throughput of the DPDK backend at packets up to 1024 bytes and rates above 200 Gbps
New: direct polling2-4 µs application send-to-receive latency in cabled loopback tests
SourceNVIDIA Technical Blog
Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

Why does an instrument need GPU-native acquisition?

NVIDIA's argument is that data movement often shapes how instruments are designed, from readout electronics to acquisition timing. If the measured data can be moved and processed fast enough, AI can adjust a process, guide an experiment or search for better operating conditions while it runs. DAQIRI is the bridge: it connects sensors optimized for measurement with programmable GPU platforms for scientific analysis, inference and feedback.

For instruments that already speak Ethernet, the post says this opens room to add processing and feedback. For new instruments, it lets designers reconsider acquisition constraints from the start. NVIDIA points to an earlier post on DAQIRI for the original design. This one focuses on a hard case.

What is the Berkeley Lab EELS camera?

EELS measures the energy electrons lose as they interact with a sample. That helps scientists identify which elements are present and how they are bonded. The measurements contain both strong and faint signals. The conventional approach captures both with two separately optimized acquisitions in rapid succession, which requires precise coordination of electron optics, exposure timing and detector readout, and then alignment and merging of the results.

The new camera from the Molecular Foundry captures strong and faint signals together. NVIDIA says that simplifies the setup and microscope optics but produces more than 500 Gbps, a stream that has to be received and processed to be useful during acquisition.

The pipeline works in four steps. FPGA readout electronics transmit detector data over Ethernet. DAQIRI receives the packets directly into GPU memory. GPU processing assembles and corrects detector frames. Then it reduces them to energy-loss spectra. NVIDIA adds that computation improves the spatial information recovered from the detector, giving submicron resolution from 5 micrometer pixels. The team can inspect data during the measurement and change acquisition parameters and regions of interest in real time, at more than 500 Gbps.

What does hardware packet reordering change?

Most networks handle out-of-order packets with protocol processing or receive-side buffering, which adds latency and memory use. DAQIRI can now hand that job to supported NVIDIA ConnectX-7 or newer NICs. For compatible sequence formats, the NIC reads the sequence field and writes each payload to its ordered place in memory, so the application receives assembled data without a software reorder copy.

The configuration is per receive queue. NVIDIA's example sets a hardware reorder engine, cyclic sequence numbering, a drop action for partial batches on timeout, delivery to GPU memory, a 42-byte header offset, 8192-byte payload slots, and a 32-bit sequence field at bit 336, with 1024 packets per output buffer. The point is that the parsing rules live in configuration, not application code. The values are for NVIDIA's example flow, so your instrument's packet layout will differ.

How much faster is the ibverbs engine?

DAQIRI adds an ibverbs engine for high-rate raw Ethernet receive and transmit, built on libibverbs and NVIDIA DevX. It is selected by default for raw Ethernet streams, and you can choose it explicitly with engine: "ibverbs" without changing how the application API is used.

NVIDIA says it delivers approximately twice the per-core throughput of the previous default DPDK backend at small packet sizes (up to 1024 bytes) and extreme rates (above 200 Gbps), based on benchmarks on DGX Spark, IGX Thor and RTX PRO server platforms.

Figure 3 in the post shows throughput nearly doubling for 64-byte and 256-byte packets versus DPDK, with similar throughput at 8 KB. That is from single-queue benchmarks on IGX Thor. Small packets are where per-packet CPU overhead dominates, which is why the gain concentrates there.

What is direct polling and why 2-4 microseconds?

Instruments that adjust settings based on GPU analysis need low-latency feedback. In direct polling mode the application thread submits and receives packets itself, which removes the handoff to a separate I/O worker. NVIDIA reports application send-to-receive latency of 2-4 µs in cabled loopback tests.

With the ibverbs engine you can mix modes: direct polling on the control queues, batched processing on the high-throughput data queues. In the configuration that means poll_mode: "direct" on the latency-sensitive transmit and receive queues. NVIDIA's benchmark command, daqiri_bench_raw_latency with a ten-thousand-sample run and a thousand warmup samples, writes results to a CSV so you can measure your own system. Replace the hardware address placeholders first.

What this means for builders and instrument makers

If you design sensors or lab software, the practical message is that acquisition and analysis can sit on the same GPU path. The post's examples are scientific, but the same pattern applies to any setup where measurements could steer a process.

For readers following NVIDIA's science and edge work, this sits beside its SETI Allen Telescope Array agent and Holoscan pulsar detection, its Nemotron chemistry models and LitmusBench and its open-source 6G research tools. On the physical-AI side, Cosmos Data Factory addresses a different bottleneck, training data, with agents.

Questions people will ask

Is this an AI model? No. DAQIRI is plumbing. It gets data to a GPU quickly and in order, so models and signal-processing code can run on it during the measurement. The "AI" in NVIDIA's headline is what you can now run on the stream: analysis, inference and feedback that adjusts the acquisition.

Why not write to disk and analyze later? Writing first means you see the result after the experiment. For a 500 Gbps stream the data volume is also large. Processing during acquisition lets the team inspect the data as it arrives and change acquisition parameters and regions of interest in real time, which the post describes as the benefit of the Berkeley Lab setup.

Why does the dual-signal camera matter? The conventional EELS approach needs two separately optimized acquisitions in rapid succession and a later merge. A camera that records strong and faint signals together removes that coordination, but only if the data pipeline can keep up. That is the pairing NVIDIA wants to show: a detector designed with GPU processing in mind.

Is the code open? NVIDIA's post points readers to the NVIDIA/daqiri repository on GitHub, the DAQIRI documentation, configuration examples and benchmarks. Check the repository for the current license and supported platforms before you plan around it.

Where does DPDK fit? DPDK was the previous default backend. The ibverbs engine replaces it as the default for raw Ethernet streams, and NVIDIA's comparison is the basis for the roughly 2x claim at small packet sizes.

Who benefits first? Teams with Ethernet-capable instruments that already stream packets. NVIDIA says those can add processing and feedback, while new instruments can be designed around GPU acquisition from the start. A sensor that only writes files over a vendor interface would need a different path.

Limits to keep in mind

  • Vendor benchmarks. The 2x throughput and 2-4 µs figures are NVIDIA's. The latency number is from cabled loopback on one host.
  • Hardware requirements. NIC reordering needs ConnectX-7 or newer. Your sensor must expose sequence fields in a compatible format.
  • Scientific claims. The submicron resolution and 500x data-rate comparison come from the NVIDIA post describing the Berkeley Lab work. We have not seen an independent paper in the sources we read.
  • What the AI does. The post covers data acquisition and GPU processing. It is not an AI model release.

How to try it

  1. Read the NVIDIA post and the DAQIRI documentation it links.
  2. Review the configuration examples for your NIC and sensor.
  3. Run the DAQIRI benchmarks to measure throughput and latency on your platform.
  4. Start with the ibverbs engine, then add reordering and direct polling for the queues that need them.

Related reading on explainx.ai:

  • NVIDIA SETI agents and the Allen Telescope Array
  • Nemotron for science and LitmusBench
  • NVIDIA Aerial and Sionna for 6G research
  • NVIDIA Cosmos Data Factory
  • NVIDIA sparse attention co-design
  • NVIDIA frontend-crates tool-calling parsers

Specifications and benchmark figures are accurate as of November 19, 2026 and come from NVIDIA.

Spotted something out of date? Let us know.
Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

  • ✓ Free
  • ✓ One email a day, no spam
  • ✓ Unsubscribe anytime
Yash Thakker

Written by

Yash Thakker

Yash is an AI expert with over 300K learners. Join his workshops →

View Yash Thakker in People in AI →

Related posts

Nov 13, 2026

NVIDIA cuDLA in Python: Run DLA Inference on Jetson Orin Without C++

NVIDIA published a cuDLA Python walkthrough on November 13, 2026 for Jetson AGX Orin. The cuda.bindings.cudla module exposes hybrid-mode DLA execution coordinated through CUDA streams, so preprocessing stays on GPU while fixed-function layers run on DLA. The same API also applies to NVIDIA RTX Spark laptops.

Nov 12, 2026

Omnii Defense: Radical Numerics Offline AI Pathogen Screening, What Wired Reports

On November 11, 2026 Wired reported that Radical Numerics unveiled Omnii Defense, an edge model that flags natural and engineered pathogens without internet access. The company ties it to Omnii-1 benchmarks from October; access stays limited to vetted testers while national labs evaluate performance.

Nov 3, 2026

AI Agents Now Run Part of the SETI Search at the Allen Telescope Array: Nexus, Holoscan and a Crab Pulsar First

NVIDIA says Breakthrough Listen and the SETI Institute now let AI agents steer the Allen Telescope Array, fix data errors and run models on a live stream of up to 2.4 terabits per second. Weeks of work shrink to hours, per the team. This post separates the vendor claims from what is independently documented.