AsHES 2020 Keynote Speaker (5:30 pm CDT)
Taisuke Boku · 2020
In the Exa-scale era, one of the most important and tough problems is how to enhance the sustained performance against the limited power budget. Traditional multi-or many-core general CPUs are still popular for easy programming and porting of general applications. However, it is getting to face to the limit by semiconductor technology limit, memory capacity per core, network bandwidth, etc. GPU represents the attached accelerator solution in heterogeneous computing thanks to its high peak performance ratio to power consumption, and moreover, recent progress on AI applications such as TensorFlow ready NVIDIA GPUs. However, GPU's extremely high performance is provided by wide width of data parallel computation both in instruction level and core/thread level which requires thousands of SIMD operations simultaneously executed. Many of success stories on GPU acceleration depend on their simple parallel execution and quite low rate of exception (if statements) handling. Another problem is the interconnection network which relies on CPU-bundle high performance network such as InfiniBand. In our research team has been focusing on the FPGA computation, which is one of the hot topics of new type of accelerators for HPC, however it is quite difficult to achieve a comparable performance with GPU especially for SIMD style applications. So, we think that a new generation of accelerated computing supported by multiple heterogeneous accelerator platform including several types of ones together on computation node. The first target is a combination of GPU and FPGA to provide 360-degree solution with SIMD and pipelined parallelism depending on the characteristics of each computation part of a large application. In this talk, I will introduce the current status of our Multi-Hetero Accelerated System running on University of Tsukuba, its hardware and software development, and real application with preliminary performance evaluation.