Managing Specialized and Heterogeneous Architectures

With the recent years of slowing of Moore/Dennard improvements designers have turned to a range of approaches for extending scaling of computer systems performance and power efficiency. One promising approach is "domain-specific architecture," which are architectures tailored to a specific problem domain, and that potentially offer significant performance gains for that domain. Examples of domain specific architectures include graphics processing units (GPUs), neural network processors used for deep learning, and processors for software-defined networks. Domain specific architectures can achieve higher performance and greater energy efficiency for four main reasons: (1) They can exploit a more efficient form of parallelism for the specific domain. (2) They can make more effective use of the memory hierarchy. (3) They can use less precision when it is adequate. (4) They can benefit from targeting programs written in domain-specific languages. Achieving significant gains through domain specific architectures will require a vertically integrated design and an
understanding of the applications, the domain-specific languages and related compiler technology, the computer architecture and organization, and the underlying implementation technology. So the purpose of the workshop will be to understand some of the specialized architectures that currently exist, and to understand some of the algorithmic optimization problems that naturally arise in these architectures. The mornings will primarily be devoted to talks by domain experts that explain various emerging specialized/heterogeneous technologies. The afternoons will be devoted to discussions of these technologies, and interesting algorithmic issues that arise in the management of these technologies.