HPC in Asia:
(A05) Poster from Japan: Accelerating HPC applications on FPGAs using OpenCL and FPGA Network
TimeWednesday, June 27th10am - 11am
DescriptionWe show two topics in this poster. One of them is OpenCL-ready high-speed 40Gbit Ethernet FPGA networking. The other is optimizing space radiative transfer code using OpenCL. We add networking functionality to the board support package (BSP) to use them from OpenCL. BSP is a hardware component to abstract differences between boards. Since the BSP provided by the board vendor does not support all peripherals on the board, we have to add controllers for Ethernet to BSP. It achieves 1μ sec latency and 4.97GB/s bandwidth (99.4% of the theoretical peak).
Accelerated Radiative transfer on grids Oct-Tree (ARGOT) is a program to solve space radiative transfer problem and has been developed in Center for Computational Sciences (CCS), University of Tsukuba. Authentic Radiation Transfer (ART) is one of algorithms used in ARGOT and is the dominant part of ARGOT program. We optimize ART algorithm to FPGAs using OpenCL. Our implementation uses channels to improve performance block RAMs in the FPGA chip. In performance comparison among FPGA, CPU and GPU, FPGA is 4.9 times faster than CPU and almost equal performance with GPU.
As future work, we will combine the network and the application to realize Accelerator in Switch (AiS) concept. It couples communications and computations tightly. We consider FPGAs can realize AiS because they can act as both of accelerators and switches.