Accessible Next-Generation storage for Efficient deep Learning pipelines

Progress in the field of artificial intelligence (AI) highly depends on data and modern hardware. The deep learning models that enable most of today’s AI applications, from speech recognition to image analysis, require large amounts of data to be trained on to reach a high accuracy. However, the processing and movement of this data in deep learning tasks is often a bottleneck and requires extensive use of hardware resources. This project aims at enabling more efficient data movement and processing in deep learning applications through leveraging the capabilities of modern storage hardware, SSDs (Solid State Drives), and open interconnect protocols, CXL (Compute Express Link). As a result, this project will reduce the total hardware needs, hence monetary costs and carbon footprint, of AI applications. Furthermore, it will allow a larger variety of AI applications, which may be impractical today due to sub-optimal use of hardware. Therefore, the project directly impacts the field of AI.

This project will bridge the gap between what the emerging storage landscape offers and what the data processing pipelines of deep learning applications need. Our objectives are to (O1) establish a methodology to measure the efficiency and sustainability of the data path in deep learning and (O2) strengthen the ease-of-use of the hardware that
facilitates fast data movement. O1 is essential to characterize the application needs and evaluate the effectiveness of proposed optimizations. O2 is essential for the wide adoption of the outcomes of this project and the capabilities of modern storage.