Tuesday, 20.6.2017, 11:30
The end of CMOS transistors scaling marks a new era for modern computer systems. As the gains from traditional general-purpose processors diminish, researchers explore the new avenues of domain-specific computing. The premise of domain-specific computing is that by co-optimizing software and specialized hardware accelerators, it is possible to achieve higher performance per power rates. In contrast to technology scaling, specialization gains are not sustainably scalable, as there is a limit to the number of ways to map a computation problem to hardware under a fixed budget. Since hardware accelerators are also implemented using CMOS transistors, the gains from specialization will also diminish in the farther future, giving rise to a "specialization wall". We explore the integration of emerging memory technologies with specialized hardware accelerators to eliminate redundant computations and essentially trade non-scaling CMOS transistors for scaling memory technology. We evaluated our new architecture for different data center workloads. Our results show that, compare to highly-optimized accelerators, we achieve an average speedup of 3.5x, save on average 43% in energy, and save on average 57% in energy-delay product.
Adi Fuchs is a PhD Candidate in the EE Department at Princeton University working on novel computer architectures. His research explores the integration of new scaling technologies with existing computer systems and domain-specific accelerators. He earned his BSc and MSc degrees, cum laude and summa cum laude, respectively, both from the EE Department at Technion - Israel Institute of Technology