Adept: Bringing Energy Efficiency Considerations to HPC

Users in the embedded software and electronic engineering fields commonly have to deal with limited power budget. HPC has always been focussed on maximising processing performance, no matter what the power cost. However, the size and processing power of the largest computing clusters now mean that the energy cost is now a very significant part of the overall system cost of the computing cluster and ways of better managing this need to be explored.

Electronic designers have long had access to tools for estimating the power consumed by a system, to ensure that it will work given a power (or temperature) budget or possibly its a battery lifetime limited system. Embedded software has fewer tools, and often the processor is assumed to be operating in one of a few power states, but some power awareness may be needed. HPC programmers typically design to meet their scientific objectives and do not pay any attention to energy required to solve their problem.

ADEPT helps raise awareness of energy efficiency in HPC, and of how software design decisions can affect the energy efficiency. Having tools to predict the energy consumption of large designs should allow the identification of any power hotspots, some of which may be suitable for offloading to more energy efficient processing solutions such as FPGAs. With accurate energy estimates, the cost of running the power hotspot code for a lifetime of work can easily be estimated, and if significantly high, justification for putting in significant effort to port the code to a higher efficiency platform can be made.  

Alpha Data hope to benefit from this increased awareness in energy efficiency in high performance computing, and continue to provide lower power acceleration solutions to the industry.