New Dimensions in Design Space and Runtime Adaptivity for Multiprocessor Systems Through Dynamic and Partial Reconfiguration: The RAMPSoC Approach

Embedded high performance computing applications have two requirements which hardly can be achieved simultaneously: high performance and low energy consumption. One solution is the exploitation of the low-level parallelism of a field programmable gate array (FPGA). Due to the manifold parameters, such as the adaptation of the clock frequency in relation to the application requirements, a better energy efficiency compared to traditional processor-based platforms can be achieved. However, the FPGA programming is time consuming until today and requires a very good understanding of the underlying hardware. There exist C-to-FPGA tools, which leverage the traditional FPGA programming using HDL-languages. However, C-to-FPGA tools can only be used for submodules and accelerators, because they do not handle the communication with the environment, e.g., camera interfaces, PCI-interfaces, etc. Furthermore, the results of an automatic code transformation are until today suboptimal in comparison to a hand coded design. Due to this fact, the interfaces have to be either programmed by hand, which is very time consuming, or they have to be bought from IP suppliers. Furthermore, these C-to-FPGA tools often have some restrictions on the input C, C++ language. In this book chapter a novel holistic approach called RAMPSoC (Runtime Adaptive Multiprocessor System-on-Chip) is presented. RAMPSoC provides a meet-in-the middle solution by combining the hardware flexibility and low power consumption of FPGAs with the software flexibility and the high-level programming paradigms of multiprocessor systems-on-chip. The RAMPSoC approach consists of a flexible and energy efficient hardware architecture, consisting of heterogeneous processing elements connected over a heterogeneous Star-Wheels Network-on-Chip, a user-guided design methodology and a new operating system for runtime resource management. RAMPSoC provides new dimensions for design space and runtime adaptivity by exploiting the features of dynamic and partial reconfiguration in FPGA-based designs. Using an object recognition algorithm, it was shown that the RAMPSoC is more energy efficient than a standard CPU and an NVIDIA Tesla GPU.