Back

AWS Trainium vs NVIDIA CUDA for Medical Image Classification: A Comprehensive Benchmark on ChestX-ray14

Fisher, G. R.

2025-12-30 radiology and imaging
10.64898/2025.12.23.25342933
Show abstract

We present a rigorous benchmark comparing AWS Trainium (trn1 instances) and NVIDIA CUDA (g5 instances with A10G GPUs) for training convolutional neural networks on medical image classification. Using the NIH ChestX-ray14 dataset with 112,120 chest radiographs and 14 thoracic disease labels, we evaluate ResNet-50 and ConvNeXt architectures across both platforms. Our key findings are threefold: (1) Trainium achieves virtually identical accuracy to CUDA for compatible architectures (ConvNeXt-Pico: F1=0.8007 vs 0.8027, {Delta}=0.25%), (2) modern CNN architectures using depthwise convolutions and LayerNorm (ConvNeXt-Tiny and larger) fail to compile or load on Trainium due to hardware constraints, and (3) Trainium is 3-5 x more expensive than CUDA for CNN training even with correct instance sizing. We document the substantial porting effort required, including four critical XLA-specific code modifications, and provide guidance for practitioners considering Trainium for computer vision workloads.

Matching journals

The top 6 journals account for 50% of the predicted probability mass.

1
Scientific Reports
based on 701 papers
Top 3%
16.1%
2
Nature Communications
based on 483 papers
Top 4%
13.3%
3
PLOS Digital Health
based on 88 papers
Top 0.7%
11.7%
4
PLOS ONE
based on 1737 papers
Top 67%
5.3%
5
Scientific Data
based on 30 papers
Top 0.4%
3.1%
6
NeuroImage
based on 36 papers
Top 1%
2.9%
50% of probability mass above
7
Nature Medicine
based on 88 papers
Top 3%
2.9%
8
npj Digital Medicine
based on 85 papers
Top 6%
2.9%
9
Informatics in Medicine Unlocked
based on 11 papers
Top 0.7%
2.4%
10
Human Brain Mapping
based on 53 papers
Top 3%
2.4%
11
Magnetic Resonance in Medicine
based on 11 papers
Top 0.7%
2.4%
12
Patterns
based on 15 papers
Top 1%
1.8%
13
JCO Clinical Cancer Informatics
based on 14 papers
Top 2%
1.6%
14
IEEE Access
based on 11 papers
Top 0.9%
1.6%
15
Imaging Neuroscience
based on 18 papers
Top 0.7%
1.6%
16
Diagnostics
based on 36 papers
Top 3%
1.4%
17
Neuro-Oncology Advances
based on 14 papers
Top 1%
1.4%
18
European Radiology
based on 11 papers
Top 2%
1.4%
19
European Heart Journal - Digital Health
based on 15 papers
Top 2%
1.3%
20
PLOS Computational Biology
based on 141 papers
Top 8%
1.3%
21
Computers in Biology and Medicine
based on 39 papers
Top 5%
1.3%
22
JMIRx Med
based on 29 papers
Top 5%
0.9%
23
npj Precision Oncology
based on 14 papers
Top 3%
0.8%
24
eBioMedicine
based on 82 papers
Top 6%
0.8%
25
The Lancet Digital Health
based on 25 papers
Top 5%
0.7%
26
Journal of Biomedical Informatics
based on 37 papers
Top 5%
0.7%