After clicking “Watch Now” you will be prompted to login or join.
High-Throughput 3D Image Reconstruction, Visualization, and Segmentation of Large-Scale Data at the Sirius Synchrotron Light Source
Thiago Vallin Spina, Brazilian Synchrotron Light Laboratory / CNPEM
GTC 2020
We'll present highly efficient tools for large-scale 3D image reconstruction, visualization, and segmentation being developed for the Sirius synchrotron light source. Sirius will be the second fourth-generation synchrotron in the world, and will acquire 3D/4D images with resolution up to <50 nm using hard X-rays. With NVIDIA, we're creating integrated pipelines to reconstruct 3D images from modalities such as coherent-diffraction imaging and transmission tomography, visualize the data in streaming mode, and segment the images to provide almost real-time feedback. We rely on multi-GPU/node CUDA programming and machine/deep learning-optimized inference to address the issues, given that each 3D image may be larger than 100 GB and may be acquired down to 1s, resulting in around 50 TB of data of a wide variety of samples (for example, biological and geological) expected to be produced every day.