Accelerating Vision-Language Models: BridgeTower on Habana Gaudi2 — x2.5 speedup with data loading

AI Impact Summary

The BridgeTower vision-language model achieves significant performance gains – up to 2.5x faster – when fine-tuned on Habana Gaudi2 hardware compared to Nvidia H100 and A100 GPUs, primarily due to hardware-accelerated data loading via Optimum Habana. This optimization leverages the Gaudi2's HPUs and their ability to efficiently handle data loading, a common bottleneck in vision model training, demonstrating a practical path to accelerate model training.