When AlphaFold2 transformed drug discovery in 2020, it depended on roughly 170,000 protein structures collected since 1971 and archived in the Protein Data Bank. Measured data remain the foundation for AI models, but modern sensors and detectors now produce data far faster than legacy “collect, store, analyze” architectures were designed to handle. Facilities such as the Linac Coherent Light Source II (LCLS-II), which generates photon pulses at a 1 MHz repetition rate, illustrate how output rates shift the bottleneck toward data handling rather than data scarcity.
NVIDIA’s DAQIRI (Data Acquisition for Integrated Real-time Instruments) moves data acquisition to a flexible, software-centric pipeline. As a high-performance networking library within the NVIDIA Holoscan Platform, DAQIRI connects high-bandwidth streaming detectors and sensors directly to the NVIDIA software ecosystem (for example, Holoscan for real-time multi-modal processing, NVIDIA TensorRT for real-time inference, and NVIDIA nvCOMP for streaming compression). DAQIRI can also stream directly into custom instrument-specific software platforms.
Why this matters
Streaming and pre-processing data at the source reduces information loss during data surges and shortens the path from measurement to discovery. DAQIRI enables instrument builders and developers to implement streaming workflows for filtering, inference, compression, event selection and adaptive control without modifying hardware. Instrument outputs can be streamed into edge supercomputing systems equipped with an NVIDIA ConnectX Network Interface Card (NIC) — ranging from NVIDIA DGX Spark and the NVIDIA IGX Platform to node and rack-based solutions such as the NVIDIA RTX Pro Server or VR200 depending on compute needs — allowing continuous experiment monitoring and real-time responses.
A-GHOST: making otherwise unsavable data searchable
The High-Luminosity Large Hadron Collider (HL-LHC) upgrade at CERN will boost luminosity by a factor of 10 over the original design. To cope, the ATLAS detector’s selection system is being upgraded: after the first stage the bandwidth of selected events rises from 100 kHz to 1 MHz, and after the second stage the events sent to storage increase from 1 kHz up to 10 kHz. Even at those rates the online system will still reject more than 99% of collisions.
The A-GHOST project uses DAQIRI to apply stronger, AI-driven searches to the data stream that the nominal selection path discards, bringing GPUs closer to raw detector outputs through efficient networking. The R&D focuses on a streaming link between planned FPGA-based boards for HL-LHC and a high-performance GPU-enabled processing farm. Led by CERN Openlab, University of Chicago and UCL scientists, the effort aims to analyze the full data stream in real time using models such as Convolutional Auto-Encoders (CAEs), temporal Convolutional Neural Networks (TCNs) and transformer-based models on prototype hardware.
How DAQIRI works under the hood
DAQIRI is built to handle high-bandwidth Ethernet traffic, including UDP and RoCE v2, at line rates of hundreds of Gbps and higher. To accomplish this, it bypasses the Linux kernel. Using the Data Plane Development Kit (DPDK), DAQIRI provides zero-copy access that routes data directly from the NIC into GPU Direct Memory Access (DMA) buffers. This kernel-bypass approach reduces latency and CPU overhead typical of conventional network stacks, ensuring massive instrument streams arrive at the GPU ready for immediate processing.
Key DAQIRI features:
- High throughput, low latency: achieve line rate on interfaces with proper hardware and CPU/NUMA tuning.
- Customized receive processing: automated packet reordering, data-type conversion and hardware-based flow steering.
- Zero memory copy to GPU: direct NIC ring-buffer access (batched and header-data split) into GPU tensors, keeping latency at PCIe transit time.
- YAML-driven configuration: optimized, customizable network boilerplate for easy deployment.
- Flexible data movement backends: support for Linux Sockets, DPDK and RoCEv2 to match application and hardware needs.
- Plug-and-play C++ and Python APIs: build a real-time application and interface with other GPU libraries quickly.
Developer experience and workflow
DAQIRI abstracts networking complexity: instrument designers supply a configuration file that defines the data path before runtime (which NIC to use, which GPU owns packet buffers, filtering rules and how payloads assemble for downstream processing). Instead of managing packets or manual memory allocations, DAQIRI batches incoming network packets directly into GPU tensors. The reorder stage can also convert payload formats (for example, int4 samples on the wire into fp16 tensors for AI inference) as part of the GPU-side assembly. Examples and templates illustrate settings such as batch sizes, timeouts (e.g., flushing partial batches after 2 ms) and buffer sizes.
With these configurations in place, application code is intentionally small: DAQIRI initializes from the YAML, returns completed bursts that contain reordered GPU-resident tensors, and the application hands those tensors directly to models or CUDA kernels. After processing, buffers are returned to DAQIRI for reuse in subsequent bursts.
Implications and use cases
DAQIRI replaces a rigid, hardware-centric collection paradigm with a software-defined, AI-enabled architecture that lets instruments process data in stream and perform real-time inference at the edge. This is especially important in high-rate scientific facilities (LCLS-II, HL-LHC/ATLAS), industrial CT and high-bandwidth software-defined radio systems, and any scenario where incoming data rates outpace traditional store-then-process workflows.
Further resources and acknowledgments
NVIDIA provides the DAQIRI GitHub repository, Getting Started documentation, benchmarks and examples for integration. Acknowledgments include David Miller and Ioannis Xiotidis (CERN) for their collaboration on using DAQIRI in the A-GHOST pipeline and their technical review, Alexis Girault for ANO documentation that informed the effort, and NVIDIA contributors Cliff Burdick, Chloe Crozier, Jay Carlson, Mahdi Azizian, Julien Jomier and the broader Holoscan team for their expertise and support.
