Accelerated X-Ray Analysis for Nanoscale Imaging (XANI) of Novel Materials

X-Ray Data Analysis Time Cut from Months to Hours

NVIDIA has demonstrated a significant acceleration in scientific computing, reducing the analysis time for massive nanoscale imaging datasets from nine months to under four hours. The breakthrough, part of the Accelerated X-ray Analysis for Nanoscale Imaging (XANI) project, processes 42 terabytes of data on a cluster of 32 NVIDIA GB200 Grace Blackwell Superchips. This addresses a critical bottleneck at X-ray free-electron laser (XFEL) facilities, enabling researchers to analyze complex materials science data in near real-time and steer experiments on the fly.

The performance improvement stems from a full-stack optimization targeting both I/O and computation. Traditional CPU-bound Python pipelines often process only a fraction of the data generated by high-repetition-rate XFEL instruments. NVIDIA's approach migrates the workflow to a GPU-centric model using a suite of new CUDA Python libraries that parallelize operations across the cluster. Key technical components of this acceleration include:

• cuPyNumeric Library: A distribution engine that partitions NumPy-like arrays across the cluster's aggregate memory, translating Python calls into parallelized tasks without requiring manual MPI coding.
• GPUDirect Storage (GDS): An I/O technology that bypasses the host CPU, allowing data to be read from high-performance storage directly into GPU memory, achieving throughput up to 700 GB/second on 16 nodes.
• Multithreaded HDF5: A custom-developed, parallelized version of the HDF5 library that overcomes the traditional single-threaded bottleneck, allowing for concurrent reads that can saturate modern storage systems.
• Batched GPU Solvers: Implementation of algorithms like Levenberg-Marquardt (JAXfit) to run nonlinear fitting for millions of detector pixels simultaneously as a single batched operation on the GPU.

This development provides a blueprint for managing exascale datasets in other scientific domains. By building high-performance libraries that integrate with the familiar Python ecosystem, the XANI project lowers the barrier for domain experts to adopt accelerated computing. This accessibility allows scientists in fields from quantum physics to materials chemistry to focus on discovery rather than HPC implementation, marking a significant step in making GPU-powered supercomputing more practical for a wider scientific community.

Strategic Takeaway: NVIDIA's acceleration of the XANI workflow demonstrates that overcoming exascale data bottlenecks requires a full-stack approach, where software innovations like cuPyNumeric and multithreaded I/O libraries are as critical as the underlying GB200 hardware for translating raw performance into practical scientific progress.