Copied


NVIDIA DAQIRI Enables Real-Time AI for Scientific Instruments

Ted Hisokawa   Nov 20, 2026 02:57 0 Min Read


NVIDIA has unveiled advancements in real-time AI capabilities for scientific instruments through its GPU-native data acquisition library, DAQIRI. This software-defined library allows high-bandwidth sensor data to be processed directly in GPU workflows without intermediate disk storage, enabling real-time analysis and feedback applications.

Revolutionizing Data Acquisition with GPU Integration

DAQIRI, part of the NVIDIA Holoscan platform, bridges high-speed sensors with programmable GPU platforms, streamlining scientific analysis and AI inference. The library’s latest release introduces features such as hardware packet reordering, improved throughput, and low-latency NIC polling for time-sensitive data processing. These enhancements are designed to optimize data movement and enable intelligent instrumentation.

Case Study: High-Dynamic-Range Electron Microscopy

A collaboration between Lawrence Berkeley National Laboratory’s Molecular Foundry and NVIDIA demonstrates DAQIRI’s potential. A newly developed high-dynamic-range electron microscopy camera captures strong and faint signals simultaneously, generating over 500 gigabits per second (Gbps) of data—more than 500 times the bandwidth of conventional electron energy-loss spectroscopy (EELS) systems. This data is transmitted via FPGA readout electronics over Ethernet, received directly into GPU memory by DAQIRI, and processed in real time.

The GPU pipeline assembles and corrects detector frames, reduces them to energy-loss spectra, and enhances spatial resolution to submicron levels. This workflow allows researchers to inspect data and adjust acquisition parameters dynamically, transforming how experiments are conducted.

Technical Enhancements in DAQIRI

The latest DAQIRI release includes:

  • Hardware Packet Reordering: Supported on NVIDIA ConnectX-7 or newer NICs, this feature offloads packet reordering tasks, reducing CPU and GPU overhead. It ensures ordered payloads are delivered directly to GPU memory for seamless processing.
  • Optimized Library Backend: Improvements in throughput and performance enhance the library’s ability to handle high-rate data streams.
  • Low-Latency NIC Polling: A direct polling mode achieves application send-to-receive latencies as low as 2-4 microseconds, critical for real-time feedback scenarios.

These advancements enable instrument makers to rethink acquisition constraints, allowing for more flexible and efficient designs.

Broader Implications for Science and Industry

DAQIRI’s integration of sensors with GPU processing has implications beyond electron microscopy. By enabling real-time data analysis and adaptive control, the library opens new possibilities for scientific exploration and industrial applications. The ability to process high-bandwidth data streams during acquisition empowers researchers and engineers to extract more actionable insights from their measurements.


Read More