Project Details
Abstract
In the deep submicron era, static power plays the major role in the total power consumption of
a chip. To the change on power consumption profile, adaptive IPC (instructions per cycle)
processors were proposed to save static energy for program execution. This project studies the
architecture and compiler design for AIPC (adaptive IPC) processors. The objective is to
achieve the energy-proportional effect: minimize the power to meet the performance
requirements. Using the MTCMOS power-gating technologies from VLSI designers, we
propose PGRF-VLIW architecture featuring power-gated and distributed register files.
Previous evaluation shows that the design makes power dissipation on functional units and
register files scale with parallelism. This project is aimed to give a proof of concepts to the
architecture and enhances the energy-proportional compilation for practical use. Research
issues are as follows.
(1) Build a precise power model for the PGRF-VLIW architecture and justify through RTL
implementation. The power model enables rigorous algorithmic research for
energy-proportional compiler optimization.
(2) Develop dynamic programming algorithm for deadline-constrained and energy-oriented
instruction scheduling. The algorithm is aimed to find an almost-optimal solution and
takes integrated consideration on the power dissipated on both functional units and
register files.
(3) Energy-aware software pipelining to realize energy-proportional computing over kernel
loops. This algorithm enables the practical use of the proposed architecture and theory.
With this project, we setup the experiment platform and theoretical foundation for
researchers from multiple disciplines to work for power-gated computing. Moreover, the
proposed technology pushes the proposed concepts from theoretical study to practical use.
Project IDs
Project ID:PB10207-1813
External Project ID:NSC102-2221-E182-028
External Project ID:NSC102-2221-E182-028
Status | Finished |
---|---|
Effective start/end date | 01/08/13 → 31/07/14 |
Fingerprint
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.