Coleman Kendrick - Using Numba for GPU acceleration of Neutron Beamline Digital Twins | SciPy 2023

"Learn how to accelerate your neutron beamline digital twins on NVIDIA CUDA GPUs using Numba, a Python library, and achieve speedups of up to 1,300 times compared to CPU-only simulations."

Key takeaways

Numba can provide significant speedups for scientific simulations on NVIDIA CUDA GPUs
It’s possible to implement a neural beamline simulation on the GPU using Numba
The speaker’s team achieved a speedup of 1,300 times using Numba compared to CPU-only simulations
Numba is easy to use for GPU acceleration with minimal code changes required
One potential challenge with using Numba is supporting conditional statements, which can be difficult to optimize for on the GPU
The speaker’s team is working on profiling and analyzing performance to optimize their Numba-based code further
The team is also planning to support other STS instruments with Numba acceleration
Numba has support for multiple platforms, including AMDs HIP architecture, but the speaker’s team did not end up using it due to lack of demand
Numba is constantly being updated and has a growing community
The speaker’s team is looking into ways to compare Numba’s performance to other packages like QPy and SciPy.

Coleman Kendrick - Using Numba for GPU acceleration of Neutron Beamline Digital Twins | SciPy 2023

More talks