Stochastic Bit-Width Approximation Using Extreme Value Theory for Customizable Processors

Abstract

Application-specific logic can be generated with a balance and mix of functional units tailored to match an application’s computational requirements. The area and power consumption of application-specific functional units, registers and memory blocks is heavily dependent on the bit-widths of operands used in computations. The actual bit-width required to store the values assigned to a variable during execution of a program will not in general match the built-in C data types with fixed sizes of 8, 16, 32 and 64 bits. Thus, precious area is wasted if the built-in data type sizes are used to declare the size of operands. A novel stochastic bit-width approximation technique is introduced to estimate the required bit-width of integer variables using Extreme Value Theory. Results are presented to demonstrate reductions in bit-widths, area and power consumption when the probability of overflow/underflow occurring is varied from 0.1 to infinitesimal levels. Our experimental results show that the stochastic bit-width approximation results in overall 32% reduction in area and overall 21% reduction in the design power consumption on a FPGA chip for nine embedded benchmarks.KeywordsPower ConsumptionTotal Power ConsumptionGumbel DistributionCustom HardwareFPGA ChipThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.