It is believed that data-stream-driven computing is power and energy efficient as compared to its counterpart, instruction-stream-driven computing. This latter requires memory access and memory control overheads while the processor is fetching task instructions from the memory. The programmer describes all the tasks as instructions in the program memory. On the other hand data-stream-driven computer is already configured or hardwired for a specific computing operation, no memory is required apart from data storage. In some contexts we refer to data-stream-driven computers as accelerators or single-purpose processors. This chapter discusses the benefit of data-stream-driven computing for better power and energy efficiency. We took matrix multiplication as an example application to compare the power and energy dissipations between load/store and non-instruction fetch-based architectures. We witnessed that single-purpose processor reduces almost 100% of the dynamic power when replacing the general-purpose processor. With the current mainstream transistor technology, morphware platforms that allow massive parallelism are the potential key for data-stream-driven computer implementations to saving energy in battery-powered embedded systems and to solve the dissipated power dilemma, as the heat becomes the bottleneck of traditional high frequency processors. If the same strategy is applied to mainstream computers and data center servers, we will not only reduce electricity bills but we will also contribute to greener computing by lowering the IT sector’s CO2 emissions.
TopComparing single-purpose and general-purpose processors for power, energy and latency has been addressed in many literatures, however we hope that further reporting may yet push the message for reaching decision makers and other involved communities.
Azzopardi, Vanderbauwhede and Moadeli (2009) raised the need of green computing systems for large data processing, they demonstrated that data-stream-driven computers on FPGAs have a potential for environment friendly computing. Their demonstration consists of comparing power and energy consumption when a documents filtering algorithm is implemented in both architectures. The algorithm consumes about 130W when employing instruction fetch-based design on a dual-core Itanium 64-bit processor running at 1.6GHz, on the other hand it consumes 1.25 watts only when employing non-instruction based design on the FPGAs platform using Virtex-4 FPGAs.
Thomas, Howes, and Luk (2009) compared the power efficiency in four different types of platform when implementing three methods of random number generations, the platforms in question are: conventional multi-core CPUs (Intel Core2), GPUs (Nvidia GTX 200), FPGAs (Xilinx Virtex-5) and MPPAs (Massively Parallel Processor Arrays) (Ambric M2000). Their finding shows that non-instruction fetch-based design on FPGAs platform has the best power efficiency with 1461.20 Msapmle/Joule (millions of samples per joule), followed by MPPAs with 150.21 MSample/Joule and GPUs with 114.55 MSample/Joule, while the multi-core CPUs led to the worst case of power efficiency with only 5.07 MSample/Joule. The figures are the average power efficiency between the three methods of random number generation, the uniform, the Gaussian and the exponential.
Single-purpose processors have been used to speed up computing in many different areas such as data mining, bioinformatics and space engineering, Baker and Parsanna (2005) employed reconfigurable technology for implementing the systolic array architecture using single-purpose processors, the implementation shown a significant performance improvement as compared to the state-of-the-art SW implementation, Harris, Jacob, Lancaster, Buhler, and Chamberlain shown that the use of application specific processors on FPGAs offered 6 to 16 as speedup factor as compared to the SW implementation on a modern general-purpose CPU, moreover Figueiredo, Stakem, Flatley and Hines (1999) stated that NASA started employing the FPGAs technology to implement single-purpose processors for efficient on board data processing to eliminate all the inefficiencies due to ground-based processing.