SHF: Small: Algorithm/Architecture Co-Design of Low Power and High Performance Linear Algebra Compute Fabrics

点击翻译按钮获取中文摘要

Until recently, the speed of a computer processors could be increasedby packing more transistors into a smaller area and increasing thefrequency. This trend is now unsustainable because of powerconstraints, with only moderate gains going forward even when puttinghundreds or thousands of traditional cores onto a chip. Thus, how toreduce power consumption while increasing performance is one of thecore concerns. It is well-accepted that specialization (designingparts of the processor for a specific task) and heterogeneity(designating different parts of the processor for different tasks) canlead to orders of magnitude improvements in both aspects. However,the question is whether such efficiency can be maintained whileproviding enough flexibility to implement a broad class of operations.Leveraging unique domain expertise, research under this projectaddresses this question for the domain of matrix computations, whichare at the core of many computational advances, both in scientifichigh-performance computing as well as in the embedded, mobile orcyber-physical domains.Observing that the largest benefits can be obtained throughspecialization at the foundations, this project is aimed atco-designing algorithms and architectures to directly realize basiclinear algebra methods in an optimized combination of hardware andsoftware. By designing a specialized Linear Algebra Processor (LAP),it is possible to achieve one to two orders of magnitude improvedefficiencies compared to traditional or proposed computerarchitectures. The questions that the project will answer, through acombination of analysis, simulation, and prototyping, include: (1) Howto best design such LAPs that can efficiently execute the full set oflinear algebra routines; and (2) How LAPs can be scaled, networkedinto clusters and integrated with application software running on oneor more host processors. The broad goal of this project is to developnovel, integrated linear algebra compute fabrics that are co-optimizedand co-designed across all layers ranging from the basic hardwarefoundations all the way to the application programming support throughstandard linear algebra software packages. This project is expectedto result in a leap in computational science and discoverycapabilities, thus enabling novel breakthroughs in industry, for theconsumer, at the national labs, in education and by scientists inacademia.