Large-scale Tomography Computation

我们提出了一种新的方法来解决记忆和计算的挑战，在体图图像重建，一个高分辨率的显微成像技术。我们的方法包括将图像梯度和衍射测量分解为块，从而减少内存占用。此外，我们提出了一种并行处理技术，可以促进大量gpu（如多达4000个gpu）之间的高效通信和并行流水线。除了GPU加速之外，我们还探索了FPGA加速器在计算机断层扫描图像重建中的使用，以提高计算和能源效率。

We present a novel approach to address the memory and computation challenges in ptychographic image reconstruction, a high-resolution microscopic imaging technique. Our method involves decomposing image gradients and diffraction measurements into tiles, resulting in a reduced memory footprint. Furthermore, we propose a parallel processing technique that facilitates efficient communication and parallel pipelining across a large number of GPUs, such as up to 4000 GPUs.In addition to GPU acceleration, we explore the use of FPGA accelerators to achieve improved computing and energy efficiency in Computed Tomography image reconstruction.