2023 Fiscal Year Research-status Report
Large-scale Tomography Computation
| Project/Area Number |
21K17750
|
|---|
| Research Institution | National Institute of Advanced Industrial Science and Technology |
|---|
Principal Investigator |
陳 鵬 国立研究開発法人産業技術総合研究所, 情報・人間工学領域, 研究員 (30890199)
|
|---|
| Project Period (FY) |
2021-04-01 – 2025-03-31
|
| Keywords | GPGPU / Tensor Core / Supercomputer / Image reconstruction / Computed Tomography / MPI |
|---|
| Outline of Annual Research Achievements |
To achieve large-scale tomographic imaging, we leverage cutting-edge technology, such as Nvidia GPUs and Tensor Cores, to tackle the challenge of producing high-resolution images. While traditional methods rely on supercomputers and are time-consuming, our goal is to achieve rapid and accurate results, even on smaller systems. Our innovative approach utilizes special GPU features, including Tensor Cores, to accelerate image generation without compromising quality. Additionally, we develop a system that enables multiple GPUs to collaborate, further enhancing image processing speed. Evaluations demonstrate that our approach significantly outperforms conventional methods, even with large datasets. In summary, our work effectively meets the demands of modern imaging technology.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
To achieve large-scale tomographic imaging, we utilize advanced Nvidia GPUs and Tensor Cores for rapid generation of high-resolution images. While traditional methods rely on supercomputers and are slow, our approach operates efficiently, even on smaller systems, such as 8 Nvidia A100 GPUs. By harnessing the power of GPUs, we expedite image generation without compromising quality. Our system facilitates collaboration among multiple GPUs, further enhancing processing speed. This proven method has garnered recognition in esteemed conferences and journals. In conclusion, our innovative use of Nvidia GPUs and Tensor Cores establishes a new benchmark for large-scale high-resolution tomographic imaging. Our solution is not only faster but also more efficient compared to traditional methods.
|
| Strategy for Future Research Activity |
Our goal is to enhance image reconstruction by optimizing computing kernels using supercomputers and various accelerators, achieving superior computational and power efficiency. We will continually leverage the latest accelerators, such as Nvidia GPUs, AMD GPUs, and FPGAs, to optimize computing kernels. Additionally, we will utilize state-of-the-art supercomputers like Fugaku and ABCI to enhance the computing capabilities for large-scale image reconstruction.
|
| Causes of Carryover |
Last fiscal year, student assistants managed image reconstruction data, while some programming engineers handled development. Due to fewer experiments, the project leader took on these tasks alone. Consequently, roles of assistants and engineers were discontinued. Meanwhile, some conferences shifted online, cutting travel expenses.
Looking ahead to the upcoming fiscal year, plans include returning to attending international conferences in person. Additionally, there are intentions to conduct more experiments with the help of research assistants and engineers.
|
-
-
-
-
-
-
-
[Journal Article] At the locus of performance: a case study in enhancing CPUs with copious 3D-stacked cache2023
Author(s)
Jens Domke, Emil Vatai, Balazs Gerofi, Yuetsu Kodama、Mohamed Wahib, Artur Podobas, Sparsh Mittal, Miquel Perics, Lingqi Zhang, Peng Chen, Aleksandr Drozd, Satoshi Matsuoka
-
Journal Title
ACM Transactions on Architecture and Code Optimization (TACO)
Volume: 1
Pages: 2767～2782
DOI
-
-
-
-
-
-
-
-
-
-
