Open

The webpage Running a Pre-trained Image Classification on IBM-AIU, demonstrates how to load and run a pre-trained model on IBM’s AIU for data prediction. This page offers insights into AIU’s performance, particularly in comparison with the GPU A100 on the DGX On-Prem cloud, helping users understand efficiency and speed differences between these hardware setups.

Performance Comparison on DGX On-Prem and IBM-AIU

Natural Language Processing (NLP) Task: Question Answering

We ran the RoBERTa model from Hugging Face 100 times across two difference hardware platforms, DGX On-Prem and IBM-AIU, to measure performance on a single question answering task (e.g., SQuAD tasks). RoBERTa, a robustly optimized BERT pretraining approach, is an extension of BERT that improves upon its performance by training with larger mini-batches, longer sequences, and dynamically changing masking patterns. Each run utilized one GPU A100 on DGX On-Prem and one IBM-AIU resource, with the model configured to handle both tokenization and fine-tuning for the SQuAD dataset.

Image Classification

We ran the ResNet50 model 100 times across two different hardware platforms, DGX On-Prem and IBM-AIU, to measure performance on a single image prediction. Each run utilized one GPU A100 on DGX On-Prem and one IBM-AIU resource. The input image was preprocessed and rescaled to a size of (3,224,224). Here’s what each dimension refers to:

• 3: number of color channels (Red, Green, Blue).

• 224: height of the image in pixels.

• 224: width of the image in pixels.

Video Classification

We ran the ResNet3D model with weights R3D_18_Weights 100 times across two different hardware platforms, DGX On-Prem and IBM-AIU, to assess performance on a single video prediction. Each run used one A100 GPU on DGX On-Prem and one IBM-AIU resource. The input video was preprocessed and rescaled to (3,16,112,112), where each dimension represents:

• 3: color channels (RGB)

• 16: frames (time steps)

• 112: height of each frame in pixels

• 112: width of each frame in pixels