The Cygnus supercomputer combines GPUs and FPGAs to provide high performance computing capabilities. It has 81 nodes with a total peak performance of 2.4 PFLOPS from GPUs, CPUs, and FPGAs. 49 nodes contain GPUs only, while 32 nodes contain GPUs, FPGAs, and high-speed interconnects between the FPGAs. The FPGAs allow for application-specific acceleration and high-speed external communication. Cygnus aims to enhance performance through mixed and variable precision operations on the FPGAs.
The Cygnus supercomputer combines GPUs and FPGAs to provide high performance computing capabilities. It has 81 nodes with a total peak performance of 2.4 PFLOPS from GPUs, CPUs, and FPGAs. 49 nodes contain GPUs only, while 32 nodes contain GPUs, FPGAs, and high-speed interconnects between the FPGAs. The FPGAs allow for application-specific acceleration and high-speed external communication. Cygnus aims to enhance performance through mixed and variable precision operations on the FPGAs.
The document describes research being conducted at the University of Tsukuba's Center for Computational Sciences. It discusses several areas of research including:
1) Developing high-performance, massively parallel numerical algorithms and libraries like FFT for simulations. This includes algorithms for saddle point problems.
2) Accelerating applications using GPUs, including developing checkpointing techniques to handle long GPU computations within time limits.
3) Research on fundamental technologies for exascale computing and beyond, including using FPGAs for computation and communication acceleration in applications.
4) Developing the ARGOT code to simulate early universe object formation using tight coupling of GPUs and FPGAs, achieving up to 17x speed
The document describes research being conducted at the University of Tsukuba's Center for Computational Sciences. It discusses several areas of research including:
1) Developing high-performance, massively parallel numerical algorithms and libraries like FFT for simulations. This includes algorithms for saddle point problems.
2) Accelerating applications using GPUs, including developing checkpointing techniques to handle long GPU computations within time limits.
3) Research on fundamental technologies for exascale computing and beyond, including using FPGAs for computation and communication acceleration in applications.
4) Developing the ARGOT code to simulate early universe object formation using tight coupling of GPUs and FPGAs, achieving up to 17x speed