Low-power and application-specific SRAM design for energy-efficient motion estimation
Author(s)Sinangil, Mahmut E. (Mahmut Ersin)
Massachusetts Institute of Technology. Dept. of Electrical Engineering and Computer Science.
Anantha P. Chandrakasan.
MetadataShow full item record
Video content is expected to account for 70% of total mobile data traffic in 2015. High efficiency video coding, in this context, is crucial for lowering the transmission and storage costs for portable electronics. However, modern video coding standards impose a large hardware complexity. Hence, energy-efficiency of these hardware blocks is becoming more critical than ever before for mobile devices. SRAMs are critical components in almost all SoCs affecting the overall energy-efficiency. This thesis focuses on algorithm and architecture development as well as low-power and application-specific SRAM design targeting motion estimation. First, a motion estimation design is considered for the next generation video standard, HEVC. Hardware cost and coding efficiency trade-offs are quantified and an optimum design choice between hardware complexity and coding efficiency is proposed. Hardware-efficient search algorithm, shared search range across CU engines and pixel pre-fetching algorithms provide 4.3x area, 56x on-chip bandwidth and 151 x off-chip bandwidth reduction. Second, a highly-parallel motion estimation design targeting ultra-low voltage operation and supporting AVC/H.264 and VC-1 standards are considered. Hardware reconfigurability along with frame and macro-block parallel processing are implemented for this engine to maximize hardware sharing between multiple standards and to meet throughput constraints. Third, in the context of low-power SRAMs, a 6T and an 8T SRAM are designed in 28nm and 45nm CMOS technologies targeting low voltage operation. The 6T design achieves operation down to 0.6V and the 8T design achieves operation down to 0.5V providing ~ 2.8x and ~ 4.8x reduction in energy/access respectively. Finally, an application-specific SRAM design targeted for motion estimation is developed. Utilizing the correlation of pixel data to reduce bit-line switching activity, this SRAM achieves up to 1.9x energy savings compared to a similar conventional 8T design. These savings demonstrate that application-specific SRAM design can introduce a new dimension and can be combined with voltage scaling to maximize energy-efficiency.
Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2012.Cataloged from PDF version of thesis.Includes bibliographical references (p. 181-189).
DepartmentMassachusetts Institute of Technology. Dept. of Electrical Engineering and Computer Science.
Massachusetts Institute of Technology
Electrical Engineering and Computer Science.