Skip to main content
Back to timeline
NVIDIA Technical BlogSource publication:

A Berkeley Lab electron microscope camera emits over 500 Gbps, and NVIDIA DAQIRI feeds it straight into GPUs for live analysis

Synopsis

A high-dynamic-range electron microscopy camera built at Lawrence Berkeley National Laboratory's Molecular Foundry merges two separately optimized acquisitions into one that captures strong and faint signals together, but produces more than 500 gigabits per second, over 500 times the conventional electron energy-loss spectroscopy data stream. NVIDIA DAQIRI, a software-defined data acquisition library, receives Ethernet packets directly into GPU memory, assembles and corrects detector frames and reduces them to energy-loss spectra during acquisition, letting the team inspect data mid-measurement and adjust acquisition parameters and regions of interest in real time.

Source-provided article image: Enabling Real-Time AI for Scientific Instruments with GPU-Native Data Acquisition
Figure 1 · NVIDIA 开发者技术博客

Interpretation

Electron energy-loss spectroscopy acquisition moves from two separately optimized exposures combined afterward to a single capture of strong and faint signals, with the camera emitting more than 500 Gbps, over 500 times the conventional data stream. The conventional approach runs two separately optimized acquisition processes in rapid succession, requiring precise coordination of electron optics, exposure timing and detector readout, then aligns and combines the resulting measurements. The Molecular Foundry camera captures strong and faint signals together, simplifying the experimental setup and microscope optics while greatly increasing the raw detector information available for processing.

FPGA readout electronics send detector data over Ethernet, DAQIRI receives the packets directly into GPU memory, and GPU processing assembles and corrects detector frames before reducing them to energy-loss spectra. DAQIRI is a software-defined data acquisition library in the NVIDIA Holoscan platform that lets developers process measurements during acquisition without first writing them to disk. Computation also improves the spatial information recovered from the detector, giving submicron spatial resolution from sensor pixels measuring 5 micrometers across, and the team inspects data during the measurement and adjusts acquisition parameters and regions of interest in real time while processing input above 500 Gbps.

Perspective

The demonstrated setting is a pipeline joining the Molecular Foundry's high-dynamic-range electron microscopy camera with DAQIRI and GPU-accelerated processing, inspecting data during measurement and adjusting acquisition parameters and regions of interest under input above 500 Gbps; throughput and latency figures come from NVIDIA benchmarks on DGX Spark, IGX Thor and RTX PRO server platforms and from cabled loopback between TX and RX ports on the same host.

Hardware reordering is defined for a specific receive queue and requires configuring how the sequence number is parsed, which reordering method to use and how missing packets are handled; the ibverbs engine's roughly twice the per-core throughput of the previous default DPDK backend is limited to small packet sizes up to 1024 bytes and extreme data rates above 200 Gbps; the 2-4 microsecond send-to-receive latency was observed in cabled loopback benchmarks.