The significance of artificial intelligence (AI) and machine learning (ML) applications has changed the landscape of computer systems: AI accelerators have started to emerge in a wide range of devices, from mobile phones to data center servers. In addition to the direct contribution of performance gain in AI/ML workloads, the introduction of AI/ML accelerators brings a new flavor of computation model, the matrix processing model, that any matrix-based algorithm can leverage in theory. However, the highly application-specific designs of these accelerators place hurdles for a wider spectrum of workloads. In this talk, Hung-Wei will discuss state-of-the-art AI/ML accelerators and share his experience in transforming existing algorithms into AI/ML-specific functions. Hung-Wei’s research group has demonstrated up to 288x speedup for database join operations by using NVIDIA’s tensor cores. compared with modern CPUs. If we can extend the design of AI/ML accelerators to support more matrix operations, a set of matrix applications, including dynamic programming-based algorithms, can achieve more than 10x speedup over conventional GPUs. Finally, Hung-Wei will talk about the new programming model, paradigms that new hardware accelerators enable, and simultaneous and heterogeneous multithreading (SHMT). By using GPUs and TPUs at the same time, SHMT reveals a 2x speedup over state-of-the-art GPU implementations. Hung-Wei will also discuss some extensions that are essential to make the upcoming revolution of general-purpose computing successful.

線上視訊請點選連結