High-Performance Reconfigurable Computing Granularity

Background

High-Performance Reconfigurable Computing (HPRC) is a computing paradigm that combines reconfigurable-based processing (e.g., FPGA technology) with general purpose computing systems, whether single general purpose processors or parallel processors. The idea is to introduce hardware flexibility to accelerate computationally intensive tasks and therefore to achieve higher performance computing compared to platforms without reconfigurable hardware. These systems not only potentially improve the performance relative to non-reconfigurable general-purpose computing systems but also reduce the energy consumption. Saves of up to four orders of magnitude in both metrics are reported for some compute intensive applications (El-Ghazawi, 2008). Almost all HPC vendors provide HPRC solutions materializing their beliefs in the capacities of reconfigurable computing as accelerators for high-performance computing.

The first commercial reconfigurable computing platform for high-performance computing was the Algotronix CHS2x4 (Algotronix) consisting of an array of 1024 processors and 8 FPGAs with 1024 programmable cells each. This architecture was followed by many other reconfigurable proposals for HPRC.

A major concern in the design of HPRC systems is how reconfigurable computing is connected to the non-reconfigurable computing side, whether general-purpose computing or dedicated computing (e.g. general purpose processors or ASICs). A few alternatives exist with different expected performances, cost and flexibility. The typical approach is to consider reconfigurable systems, generally FPGAs, mounted in a board that is connected to the main system using some serial bus to operate as a coprocessor. The approach has a relative low cost but has a severe limitation from the serial communication between the host and the coprocessors whose bandwidth determines that for the architecture to be computationally efficient the computation to I/O ratio must be high. To improve the co-processing solution, a few architectures have implemented direct point-to-point connections among the co-processors. This speeds-up the communications between reconfigurable co-processors but keeps the communication bottleneck between the host system and the co-processing system. Some works have refined the communication between the host and the FPGA co-processors through a dedicated network interface. A well-known example of this architecture is Cray XD1 (Cray Inc., 2006). To speed-up even more the communication between the host CPU and the reconfigurable units, all units can be connected to a single communication network, like a shared memory system. In this case, all processing units see each other as part of a unique architecture with access to shared memories (SGI RASC RC100 blade from Silicon Graphics).