GPU-Native Data Acquisition Unlocks Real-Time AI for Scientific Instruments

NVIDIA DAQIRI enables real-time AI in scientific instruments by moving high-bandwidth sensor data directly to GPUs. This accelerates analysis, simplifies…

Scientific and industrial settings increasingly leverage artificial intelligence to analyze sensor measurements, guiding experiments and optimizing processes in real time. However, the sheer volume and velocity of data generated by advanced sensors often outpace traditional data acquisition and processing capabilities. NVIDIA DAQIRI, a software-defined library within the NVIDIA Holoscan platform, addresses this by enabling GPU-native data acquisition. It allows for high-bandwidth instrument and sensor data to be moved directly into GPU memory for processing during acquisition, eliminating the need for interim storage and providing immediate insights. This advancement offers instrument designers greater flexibility and opens new possibilities for real-time feedback applications.

Real-Time Electron Microscopy with GPU Acceleration

Lawrence Berkeley National Laboratory’s Molecular Foundry illustrates the transformative potential of integrating advanced detectors with GPU processing. Researchers there developed a high-dynamic-range electron microscopy camera capable of simultaneously capturing both strong and faint signals. While this simplifies the experimental setup for techniques like electron energy-loss spectroscopy (EELS), it generates an immense data stream exceeding 500 gigabits per second (Gbps)—over 500 times faster than conventional EELS data. To make this data useful in real time, a specialized acquisition and processing pipeline is essential. Leveraging DAQIRI, detector packets are received directly into GPU memory, allowing for immediate assembly and correction of frames, and reduction to energy-loss spectra. This GPU-accelerated workflow also enhances spatial information, enabling submicron resolution from pixels measuring 5 micrometers across. The collaboration enables researchers to inspect data and adjust acquisition parameters or regions of interest during the measurement, fundamentally altering the pace and interactivity of scientific discovery.

Cutting-Edge DAQIRI Enhancements for Efficiency

The latest release of DAQIRI introduces several critical features designed to optimize data movement and processing for intelligent instruments:

Optimized Data Handling and Low-Latency Communication

First, DAQIRI now supports hardware packet reordering on NVIDIA ConnectX-7 or newer network interface cards (NICs). This capability offloads the reordering task from the CPU and GPU, preserving their resources for computational analysis. The NIC parses sequence fields and places payloads directly into their correct memory positions, delivering ordered data to applications without requiring a separate software reordering step, which reduces both latency and memory overhead.

Second, an optimized ibverbs engine has been introduced for efficient raw Ethernet packet handling. This engine, built on libibverbs and NVIDIA DevX, is now the default for raw Ethernet streams and significantly improves throughput. Benchmarks on NVIDIA DGX Spark, IGX Thor, and RTX PRO server platforms show the ibverbs engine nearly doubles per-core throughput for small packet sizes (up to 1024 bytes) at extreme data rates exceeding 200 Gbps, compared to the previous DPDK backend.

Finally, for applications requiring rapid feedback, DAQIRI offers a direct polling mode. This mode reduces communication latency by allowing the application thread to directly receive and submit packets, bypassing an intermediate I/O worker. In cabled loopback tests, this has demonstrated application send-to-receive latencies as low as 2-4 microseconds. This feature is particularly valuable for control queues where timely GPU analysis must trigger immediate adjustments to experimental or instrument settings.

Practical Implications and Outlook

The integration of GPU-native data acquisition with real-time AI holds profound implications for scientific research and industrial automation. It allows instrument makers to design around critical measurements and desired responses, rather than being constrained by data bottlenecks. For researchers, it means extracting more comprehensive information from experiments, faster iteration cycles, and the ability to make informed decisions dynamically. While these advancements rely on specific NVIDIA hardware and software ecosystems, the ability to process high-bandwidth data streams directly on GPUs during acquisition marks a significant leap forward in connecting advanced sensors with powerful computational platforms. Researchers interested in these capabilities are encouraged to explore the DAQIRI documentation and configuration examples, and to benchmark its performance on their own hardware to understand its full potential for enabling real-time scientific insight.

Source and news date

Enabling Real-Time AI for Scientific Instruments with GPU-Native Data Acquisition — 2026-11-19

Related ServerNet services

ServerNet Support

ServerNet engineering & editorial team — specialists in infrastructure, networking and web hosting.

Linux Hosting
Share:

Comments 0

No comments yet — be the first!

Leave a comment

Related service

Linux Hosting

PHP & MySQL hosting on NVMe RAID-10 with LiteSpeed — the solid base for any website, from personal blogs to enterprise Laravel apps. At a price competitors can't explain.