AI · Nov 9, 2026
OpenAI’s Control Room puts a fail-closed kill switch inside the inference engineWhen the detector outruns the disk: Nvidia's DAQIRI and a 500-gigabit microscope
· Updated · Signal Desk newsroom
Analysis of reporting by NVIDIA Developer. How we work

A Berkeley Lab electron microscopy camera streams more than 500 Gbps straight into GPU memory, letting scientists adjust experiments mid-measurement. Nvidia's DAQIRI update makes the case that instruments should be designed around GPUs from the start.
Key takeaways
- DAQIRI, a Holoscan library, moves instrument data over Ethernet directly into GPU memory so it can be processed during acquisition without first hitting disk.
- A high-dynamic-range EELS camera from Lawrence Berkeley National Laboratory's Molecular Foundry produces more than 500 Gbps, which Nvidia says is over 500x conventional EELS streams.
- GPU processing assembles frames, produces spectra and recovers submicron spatial resolution from 5-micrometer pixels, with real-time adjustment of acquisition settings.
- New features include hardware packet reordering on ConnectX-7 or newer NICs, an ibverbs engine with about twice the per-core throughput of the old DPDK backend in some conditions, and a direct polling mode with 2-4 microsecond loopback latency.
- The performance figures are Nvidia's own benchmarks, and the stack ties instrument design closely to Nvidia hardware.
Instruments have been designed around data bottlenecks
A lot of scientific hardware is shaped not by what scientists want to measure but by how fast the resulting data can be moved and processed. Readout electronics, acquisition timing and even experimental procedure bend around that limit. Nvidia's latest DAQIRI post asks the obvious follow-up: what could instrument makers simplify if processing kept pace with the sensor?
DAQIRI is a software-defined data acquisition library within Nvidia's Holoscan platform. It takes high-bandwidth sensor data arriving over Ethernet and places it in GPU-ready form, so analysis, AI inference and feedback can happen while the measurement is still running rather than after the data is written to disk. Existing Ethernet instruments can bolt this on; new instruments can be designed with it in mind.
The microscope that makes the case
The centerpiece is a high-dynamic-range camera for electron energy-loss spectroscopy, or EELS, developed at Lawrence Berkeley National Laboratory's Molecular Foundry. EELS measures how much energy electrons lose passing through a sample, revealing which elements are present and how they bond. The signals vary enormously in strength, so the conventional approach runs two separately optimized acquisitions in quick succession and then aligns and merges them, which demands careful coordination of optics, exposure and readout.
The new camera captures strong and faint signals together, simplifying the optics and setup, but at the cost of a data stream above 500 gigabits per second, which Nvidia says is more than 500 times a conventional EELS stream. In the Berkeley-Nvidia setup, FPGA readout electronics send packets over Ethernet, DAQIRI lands them directly in GPU memory, and GPU code assembles and corrects frames and reduces them to spectra. Computation also recovers submicron spatial resolution from 5-micrometer pixels. Researchers can inspect data during the measurement and change acquisition parameters and regions of interest on the fly.
This is the strongest kind of example: hardware that only makes sense because the processing exists. It echoes Nvidia's broader SC26 theme of putting GPUs and agents in the loop of science, which we covered in Ian Buck's SC26 keynote.
Under the hood, and the fine print
The new DAQIRI release adds three pieces. Packet reordering can now be offloaded to ConnectX-7 or newer NICs, which parse sequence numbers and place payloads in order in memory, avoiding a software copy. A new ibverbs engine, now the default for raw Ethernet streams, delivers roughly twice the per-core throughput of the previous DPDK backend at small packet sizes up to 1,024 bytes and data rates above 200 Gbps, based on tests on DGX Spark, IGX Thor and RTX PRO server platforms. And a direct polling mode lets the application thread send and receive packets itself; Nvidia reports send-to-receive latency of about 2 to 4 microseconds in a cabled loopback test.
Those numbers are vendor benchmarks under favorable conditions, and loopback latency is a best case, not the latency of a real control loop through an instrument. The bigger strategic point is lock-in. Hardware reordering requires Nvidia NICs, the engine leans on Nvidia's DevX, and the processing runs on Nvidia GPUs. For a national lab building a one-off detector, that is a fair trade for capability. For commercial instrument makers deciding on a platform for a decade, it is a dependency worth weighing. The same pattern runs through Nvidia's CUDA-X push into science.
Who feels it
- Instrument makers
- GPU-native acquisition can simplify optics and electronics, but ties designs to Nvidia networking and GPUs.
- Microscopy and materials scientists
- Real-time spectra and steering mean less wasted beam time and fewer repeated runs.
- Lab IT teams
- 500 Gbps streams demand high-end NICs, GPU servers and new operational skills.
- AI-for-science developers
- Live data opens room for models that guide experiments as they run.
What to watch
- Peer-reviewed results from the Molecular Foundry camera using GPU processing.
- Commercial instrument vendors announcing DAQIRI-based products.
- Independent latency measurements in real feedback loops, not loopback tests.
- Whether DAQIRI support extends to non-Nvidia NICs or open alternatives emerge.
FAQ
What is NVIDIA DAQIRI?
A software-defined data acquisition library in Nvidia's Holoscan platform that moves high-bandwidth instrument data directly into GPU memory for processing during acquisition.
How fast is the Berkeley Lab EELS camera?
It generates more than 500 gigabits per second, which Nvidia says is over 500 times faster than conventional EELS data streams.
What is new in the latest DAQIRI release?
Hardware packet reordering on ConnectX-7 or newer NICs, an ibverbs engine with higher per-core throughput, and a direct polling mode for low-latency feedback.
More on AI
AI · Nov 11, 2026
OpenAI’s $2 million for children’s AI safety research matters for one promise: publishing bad newsAI · Nov 17, 2026
OpenAI says its models now shave area off production chips, including AMD’s next flagshipsAI · Nov 5, 2026
NVIDIA’s four open Cosmos 3 Nano models let robot and driving teams simulate worlds that reactAI · Nov 3, 2026
OpenAI’s GPT-5.6 Cyber reaches AWS GovCloud through Daybreak Red, with two approval gates