Box Muller (BM) algorithm is extensively used for generation of high quality Gaussian Random Numbers (GRNs) in hardware. Most efficient published implementation of BM method utilizes transformation of 32-bit data path to 16 bits and use of first degree piece-wise polynomial approximation to compute logarithmic and square root functions. In this work, we have performed extensive error analysis to show that coefficient memory for polynomial approximation can be reduced by more than 35 percent without compromising on quality of generated Gaussian samples. This also reduces complexity of corresponding address generator, which requires most hardware resources. We have also used more efficient and statistically accurate skipahead Linear Feedback Shift Registers to generate uniformly distributed numbers for the BM algorithm. Complete hardware implementation utilizes only 407 slices, 03 DSP blocks and 1.5 memory blocks on Xilinx Virtex-4 XC4VLX15 operating at 230 MHz while providing a tail accuracy of 6.6σ. This is better in terms of accuracy and hardware utilization than any of the previously reported architecture.
Abstract-An efficient hardware implementation of Gaussian Random Number (GRN) generator based on Central Limit Theorem (CLT) is presented. CLT, although very simple to implement, is never used to generate high quality Gaussian numbers. This is due to the fact that direct implementation of CLT provides very poor accuracy in tail regions of the probability density function. In this work, we have shown that it is possible to achieve high tail accuracy by empirically computing the error in CLT, which can be compensated with a simple correction algorithm. The error has been modeled as first degree piece-wise polynomial approximation, using a novel non-uniform segmentation algorithm to compute the coefficients of polynomial segments. A novel hardware architecture of GRN generator is presented which requires only 420 slices and 1 DSP block of Xilinx Virtex-4 XC4VLX15 operating at 220 MHz. This resource utilization is better than any of the previously reported designs. Demonstrated for the tail accuracy of 6σ, the GRN generator design is scalable to achieve even higher accuracy with minimal increase in hardware resources. The accuracy of GRN generator is validated using statistical goodness of fit tests.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.