AWS Trainium vs. Nvidia CUDA for Medical Image Classification (www.medrxiv.org)

A recent benchmark study has compared AWS Trainium and NVIDIA CUDA's performance for training convolutional neural networks (CNNs) in medical image classification, utilizing the NIH ChestX-ray14 dataset. The research evaluated ConvNeXt and ResNet-50 architectures to measure effectiveness, revealing that Trainium delivers comparable accuracy to CUDA for several compatible models, with minor F1 score differences (ConvNeXt-Pico: F1=0.8007 vs. CUDA's 0.8027). However, the study identified limitations, as larger CNN architectures employing advanced techniques like depthwise convolutions and LayerNorm could not be compiled on Trainium due to hardware constraints. Significantly, the analysis highlights that using Trainium can be 3–5 times more expensive than CUDA for CNN training, even with properly sized instances, and necessitates substantial code adjustments, including four critical modifications for XLA compatibility. This study serves as a critical resource for AI/ML practitioners, providing valuable insights into the practicalities and challenges of utilizing Trainium for computer vision applications, thereby shaping decisions on infrastructure choices in the AI community.