Home > Storage > PowerScale (Isilon) > Industry Solutions and Verticals > Analytics > Power AI with Dell and NVIDIA > ResNet-50
ResNet-50 is a real-world image classification dataset that has become a standard benchmark to characterize the performance of a deep learning training workflow on storage and GPU compute platforms. This benchmark performs training of an image classification convolutional neural network (CNN) on labeled images using MXNet. Essentially, the system learns whether an image contains a cat, dog, car, train, and so on. The well-known ILSVRC2012 image dataset (often referred to as ImageNet) was used. This dataset contains 1,281,167 training images in 144.8 GB1. All images are grouped into 1000 categories or classes.
The individual JPEG images in the ImageNet dataset were converted to RecordIO format. The dataset was not resized, not normalized, and no preprocessing was performed on the raw ImageNet JPEG images. It maintains the image compression offered by the JPEG format. The total size of the dataset remained roughly the same (148 GB). The average image size was 115 KB.
The ResNet-50 results show clear scaling from one to four DGX systems, thus reducing the training time as more DGX H100 systems and GPUs are engaged. The GPU Utilization line demonstrates the ability of the PowerScale cluster to sufficiently maintain a >90% utilization across all GPUs across all systems throughout the epochs.