Paper Info
Title
Building Scalable and Highly Efficient Accelerators Near the End of Conventional Scaling
Abstract
The dark silicon concept is around for 10 years now in which we had to get used to diminished returns after each scaling step while the sharply increasing complexity of manufacturing processes decelerated scaling and raised costs dramatically. The combination of all those factors lead to a long-awaited and slowly emerging (temporary?) slow-down of Moore's Law. There have been many reactions to and many projections based on this development but were they accurate, and how will we be able to create highly efficient accelerators in the future while scaling their performance? What are the problems when either targeting the highest energy-efficiency or highest performance? How did dark silicon age? In this paper, the author approaches these questions based on observations in generations of technology and experiments in cutting-edge technologies. One cornerstone of the observations are the design experiences of the recently announced MN-Core deep learning training accelerator which achieves more than 1 TFlop/W in half-precision.
Year
DOI
Venue
2019
10.1109/MCSoC.2019.00014
2019 IEEE 13th International Symposium on Embedded Multicore/Many-core Systems-on-Chip (MCSoC)
Keywords
DocType
ISBN
low-power,scaling,accelerators
Conference
978-1-7281-4883-0
Citations 
PageRank 
References 
0
0.34
6
Authors
1
Name
Order
Citations
PageRank
Johannes Maximilian Kühn101.35