Niteya Shah
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Niteya Shah has a total of 4 co-authors. Here are the top co-authors by number of shared publications:
Beau Johnston
1 shared publication- Beau Johnston
Oak Ridge National Laboratory, Oak Ridge, Tennessee, USA
1 shared publication Wu-Chun Feng
1 shared publication- Wu-Chun Feng
Virginia Tech, Blacksburg, USA
1 shared publication
Latest Publications
- Open Access
- 24 October 2026
- Published by Elsevier in SSRN Electronic Journal
On the Efficacy of PyTorch for Integrated High-Performance Computing and Machine Learning: A Case Study in Computational Physics
Python has become a popular language for scientific computing due to its productivity and ease of use, with frameworks such as PyTorch delivering portable performance without exposing users to the complexity of compiled programming models. While PyTorch is well established in machine learning (ML), its suitability for scientific applications that incorporate ML remains unclear.In this work, we evaluate PyTorch as a programming model for high-performance scientific computing using a differentiable sampling algorithm from computational physics embedded within an end-to-end training workflow. We compare PyTorch against C++ implementations using OpenMP and SYCL across CPUs and GPUs. In addition, we evaluate Triton as a lower-level optimization path within the PyTorch ecosystem. We analyze forward and reverse performance, scaling, and programmer productivity to quantify the trade-offs between different backends in a Python-based workflow.Our results show that conventional HPC implementations require (7.7x)--(11.8x) more source lines of code than PyTorch, while PyTorch still reaches about 80\% of optimized C++/OpenMP performance on the CPU. On GPUs, SYCL outperforms PyTorch by up to (2.7x), but Triton closes much of this gap, performing within 7% of the fastest implementation across all three accelerators. These findings highlight both the strengths and limitations of PyTorch for scientific workflows that incorporate machine learning.
- 24 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessOn the Efficacy of PyTorch for Integrated High-Performance Computing and Machine Learning: A Case Study in Computational Physics
Python has become a popular language for scientific computing due to its productivity and ease of use, with frameworks such as PyTorch delivering portable performance without exposing users to the complexity of compiled programming models. While PyTorch is well established in machine learning (ML), its suitability for scientific applications that incorporate ML remains unclear.In this work, we evaluate PyTorch as a programming model for high-performance scientific computing using a differentiable sampling algorithm from computational physics embedded within an end-to-end training workflow. We compare PyTorch against C++ implementations using OpenMP and SYCL across CPUs and GPUs. In addition, we evaluate Triton as a lower-level optimization path within the PyTorch ecosystem. We analyze forward and reverse performance, scaling, and programmer productivity to quantify the trade-offs between different backends in a Python-based workflow.Our results show that conventional HPC implementations require (7.7x)--(11.8x) more source lines of code than PyTorch, while PyTorch still reaches about 80\% of optimized C++/OpenMP performance on the CPU. On GPUs, SYCL outperforms PyTorch by up to (2.7x), but Triton closes much of this gap, performing within 7% of the fastest implementation across all three accelerators. These findings highlight both the strengths and limitations of PyTorch for scientific workflows that incorporate machine learning.
- Open Access
- 19 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 309-317
On the Efficacy of PyTorch for High-Performance Computing: A Case Study in Computational Physics
- 19 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 309-317
Open AccessOn the Efficacy of PyTorch for High-Performance Computing: A Case Study in Computational Physics