Jungfraujoch: kilohertz data acquisition and on-the-fly image analysis for macromolecular crystallography beamlines

• Leonarski Filip (PSI)

Room: tba Paul Scherrer Institute is building a 9 Mpixel JUNGFRAU detector for macromolecular crystallography (MX) applications [1] operating at frame rates up to 2 kHz. This detector will produce a stream of 38 GB/s raw, uncompressed data. Collecting and analyzing X-ray diffraction images at such high data rates is a significant computing challenge [2]. To overcome this challenge, we have developed Jungfraujoch, a read-out system, that can handle up to 50 GB/s data rate within a single standard (2U) sized server. The main building block of this system is an FPGA board, operating as a smart network card. Besides functionality to receive network packet, the card also implements conversion of JUNGFRAU raw data to photons, initial data analysis (e.g., spot finding), and a bitshuffle precompression filter. The FPGA design is accompanied by a software layer for compression, streaming, and HDF5 writing In the talk, I will present the architecture and performance of the Jungfraujoch system, accompanied by experimental results from protein crystallography beamtimes. 1. F. Leonarski, S. Redford, A. Mozzanica et al. Nat. Methods 15, 799–804 (2018). 2. F. Leonarski, A. Mozzanica, M. Brückner et al. Struct. Dyn. 7, 014305 (2020).

Azimuthal integration of area-detector data on field-programmable gate arrays for data reduction

• Zdeněk Matěj

Room: tba Hardware computing accelerators as graphical processing units (GPUs) are already extensively used for data analysis from photon science experiments. Nowadays even quantum computers are available in commercial clouds. We have started to explore a potential of Field-Programmable Gate Arrays (FPGAs) as computing accelerators for data reduction at MAX IV synchrotron Laboratory. The use of FPGAs for data reduction is demonstrated here on a task of Azimuthal Integration (AZINT) of area-detector data from powder diffraction and small angle scattering. Beside these two most known application cases the underlying procedure i.e. bin-counting has many other applications, in particular in background estimation or image analysis. The new potential of FPGAs for complex data analysis originates from recent progress in tools allowing scientific software developers to easily program FPGAs, prototype and implement algorithms on them with complexity fitting scientific requirements. It is demonstrated that AZINT can process 600 Gb/s of uncompressed data stream, i.e. about 20-40 Gpixels/s, on a single commercial FPGA available at photon and neutron facilities or in public clouds. FPGAs are usually more energy-efficient in comparison to CPUs and GPUs. However, their main advantage may be that beside high throughput they allow data reduction and filtering with well-defined and low latencies. This enables experiments with x-rays as a real-time probe. In this contribution we explain details of the AZINT/bincount algorithm [1,2], tools we are using to implement flexible solutions and discuss how we would like to use these compute accelerators for data reduction at MAX IV. 1. https://github.com/maxiv-science/azint 2. https://gitlab.com/MAXIV-SCISW/compute-fpgas/bincount