Data-transformation algorithms for minimizing bit flips on GPU data buses
Excessive power consumption and cooling costs in computing centers as well as limited battery life in mobile devices make energy optimization an important area of research. In CMOS technology, dynamic energy is primarily consumed when switching from one state to another. In particular, the charging and discharging of long wires consume a significant amount of energy. GPU-based accelerators are widely used to solve many complex data-intensive problems, which tend to transfer large amounts of data to/from main memory. Therefore, GPU data buses contribute a significant portion of the total energy expenditure. As a consequence, encoding the data to minimize bit flips has the potential to greatly reduce the amount of energy consumed by data buses. The existing commercially available solution to reduce bit flips is called Data Bus Inversion (DBI). This thesis introduces more effective bit-flip minimization algorithms, which can eliminate about 9% more bit flips than DBI.
Bit-flips, Data-transformation algorithms
Nagarajan, A. (2017). <i>Data-transformation algorithms for minimizing bit flips on GPU data buses</i> (Unpublished thesis). Texas State University, San Marcos, Texas.