GPU optimization techniques to accelerate optiGAN-a particle simulation GAN

Area of Science:

• Artificial Intelligence
• High-Performance Computing
• Computational Physics

Background:

• The complexity of AI models necessitates specialized hardware like Graphics Processing Units (GPUs) for efficient training.
• Despite hardware advancements, a gap persists between computational demands and current GPU capacity.
• Software optimization is essential to maximize hardware utilization and bridge this performance gap.

Purpose of the Study:

• To present and analyze general GPU optimization techniques for efficient AI model training.
• To demonstrate the application of these optimizations on the optiGAN model for generating optical photon distributions.
• To evaluate the performance improvements achieved through these software optimizations.

Main Methods:

• Implementation of general GPU optimization techniques during the training of the optiGAN model.
• Utilizing an 8GB Nvidia Quadro RTX 4000 GPU for training and performance analysis.
• Employing the Nvidia Nsight Systems profiler to measure execution time and memory consumption.

Main Results:

• The applied GPU optimizations resulted in an approximate 4.5x increase in runtime performance compared to naive training.
• Performance was evaluated based on execution time and memory usage, demonstrating significant efficiency gains.
• Model performance was maintained without compromise despite the accelerated training.

Conclusions:

• Software optimization techniques can substantially enhance GPU utilization for training complex AI models like optiGAN.
• The developed optimization strategies offer a practical approach to accelerate the training of generative adversarial networks.
• Future work will focus on scaling the optiGAN model across multiple GPUs for even greater computational efficiency.