Challenges in Transition

Abstract

Modern emerging workloads such as analytics, graph, and deep learning, rapidly appear. These are written by non-Ninja programmers. Modern hardware platforms are becoming complex due to deployments of hardware accelerators such as GPGPU and FPGA. It is not easy for them to fully exploit these capabilities. Our recent challenges are to achieve high performance of these workloads. In this talk, I will review how hardware platform, workload, and software for high performance computation were changing. I will then think about what are better approaches for users to describe their problems with high productivity and performance. I will talk about technical approaches and challenges in these descriptions to exploit hardware capabilities. We need to think what information we should get and what optimizations we can do for future hardware system.