Optimal digital system design in deep submicron technology

• Main Author: Heo, Seongmoo, 1977-
• Other Authors: Krste Asanović.
• Format: Others
• Language: English
• Published: Massachusetts Institute of Technology 2007
• Subjects: Electrical Engineering and Computer Science.
• Online Access: http://hdl.handle.net/1721.1/36135

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2006. === Includes bibliographical references (p. 165-174). === The optimization of a digital system in deep submicron technology should be done with two basic principles: energy waste reduction and energy-delay tradeoff. Increased energy resources obtained through energy waste reduction are utilized through energy-delay tradeoffs. The previous practice of obliviously pursuing performance has led to the rapid increase in energy consumption. While energy waste due to unnecessary switching could be reduced with small increases in logic complexity, leakage energy waste still remains as a major design challenge. We find that fine-grain dynamic leakage reduction (FG-DLR), turning off small subblocks for short idle intervals, is the key for successful leakage energy saving. We introduce an FG-DLR circuit technique, Leakage Biasing, which uses leakage currents themselves to bias the circuit into the minimum leakage state, and apply it to primary SRAM arrays for bitline leakage reduction (Leakage-Biased Bitlines) and to domino logic (Leakage-Biased Domino). We also introduce another FG-DLR circuit technique, Dynamic Resizing, which dynamically downsizes transistors on idle paths while maintaining the performance along active critical paths, and apply it to static CMOS circuits. === (cont.) We show that significant energy reduction can be achieved at the same computation throughput and communication bandwidth by pipelining logic gates and wires. We find that energy saved by pipelining datapaths is eventually limited by latch energy overhead, leading to a power-optimal pipelining. Structuring global wires into on-chip networks provides a better environment for pipelining and leakage energy saving. We show that the energy-efficiency increase through replacement with dynamically packet-routed networks is bounded by router energy overhead. Finally, we provide a way of relaxing the peak power constraint. We evaluate the use of Activity Migration (AM) for hot spot removal. AM spreads heat by transporting computation to a different location on the die. We show that AM can be used either to increase the power that can be dissipated by a given package, or to lower the operating temperature and hence the operating energy. === by Seongmoo Heo. === Ph.D.