Previous design efforts have shown that high performance, power efficient processing solutions exist if the hardware and target application are carefully co-optimized. The customized hardware in widespread use for image compression, cellular modems, etc., provides dramatic improvements in compute efficiency over general-purpose systems and demonstrates the advantages of this approach. Unfortunately the enormous engineering costs of designing these solutions make this approach out of reach for all but the largest markets, making it very difficult for the DoD to create the efficient computing systems it needs.
The remaining Specialized, Extremely-Efficient Computing (SEEC) effort of the Ubiquitous High Performance Computing (UHPC) Program will create tools and design methods that can encapsulate designer knowledge. That is, rather than recording the specific end result a designer manually creates for a design challenge, the effort will record the methods used to determine optimal (or good) solutions, and how these solutions are transformed to software and hardware. By capturing this meta-level design information, it can be used to automate some of the design process to reduce the overall effort required to develop application specific solutions. The goal is to create an infrastructure for creating domain specific languages, while still allowing for application specific performance and energy optimizations. This domain specific generation concept will be pursued to support the design of application customized hardware as well as the applications that run on it, creating an application specific hardware generator.
Dr. Joseph Crossjoseph.email@example.com