GPU Acceleration for Data Processing and Analytics

Graphics processing units, or GPUs, are widely employed as hardware accelerators in various applications, such as algorithmic trading, computer vision, and large language model training. In particular, NVIDIA’s GPUs, together with its Compute-Unified Device Architecture (CUDA) interface, provide a massively parallel platform for general-purpose computing. However, it is often challenging to accelerate data processing and analytical tasks on the GPU when they are irregular and do not match well the GPU architecture or programming paradigm. In this talk, I will discuss general methodologies as well as specific design and implementation techniques on using the GPU to accelerate such tasks, and compare them with CPU-based solutions. With the prevalence of GPU-equipped computing resources and big data applications of increasing complexity and scale, more opportunities and challenges will arise in this space.