As the performance improvements of general-purpose processors or even GPUs fall behind the rapid growth of demands in AI/ML and reality applications, modern computers rely on hardware accelerators that use specialized circuits for a target application domain to support the desired user experience. However, applications besides these hardware-accelerated domains still suffer from the retarded performance improvements. Theoretically, these accelerators can benefit a broader spectrum of applications since their circuits implement mathematical functions or simulate physical phenomena that many problems can leverage in their solutions. However, as these accelerators typically simplify the supported functions for their target domains, using them for other applications becomes challenging. 

 

Hung-Wei will share his recent experiences using and extending modern AI/ML and ray tracing accelerators for the most critical tensor linear algebra problems in this talk. Hung-Wei will discuss using low-precision matrix multipliers that AI/ML accelerators provide to perform standard precision matrix operations that general-purpose linear algebra needs. Hung-Wei will also present the use of ray-tracing hardware to address the tree search problems and sparse matrix problems.