Speaker
Description
Abstract
Ultra-fast time-resolved synchrotron computed tomography (CT) can acquire up to 1000 tomographs per second, enabling the study of highly dynamic processes but producing data streams that challenge transmission, reconstruction, analysis, and storage. To support smart experiments and rapid feedback at the TOMCAT beamlines of the Swiss Light Source, we developed a live reconstruction pipeline that sustains detector-rate processing in real time.
The system integrates high-speed acquisition, remote direct memory access (RDMA) streaming, and multi-GPU reconstruction. On the data acquisition node, 12-bit GigaFRoST camera output is converted to 16-bit format, grouped into 3D chunks, and stored in shared-memory buffers. Metadata are exchanged with ZeroMQ, while image payloads are transferred with Unified Communication X (UCX). On the reconstruction node, data are buffered in host memory, partitioned into chunks, and processed on four NVIDIA H100 GPUs. Host-to-device and device-to-host transfers are overlapped with computation through CUDA streams, and the reconstruction workflow is implemented on top of Tomocupy.
The pipeline achieves up to 15 GB/s reconstruction throughput in 16-bit format, matching the full detector streaming rate of 7.7 GB/s at 12-bit output. This enables continuous reconstruction with sufficiently low latency for downstream analysis and online decision-making. The resulting capability provides the basis for adaptive data acquisition strategies that can suppress redundant data, preserve critical events, and feed reconstructed information back to the beamline control system. These results show that live reconstruction is a practical enabling technology for smart ultra-fast synchrotron CT and is well aligned with the performance requirements of the Swiss Light Source 2.0 upgrade.
Acknowledgements
This project is funded by the Swiss Data Science Center (SDSC), No. C23-02L.
