Dynamic bitwidth assignment for efficient dot products
作者： Simon Joel Schmidt , David Boland
DOI: 10.23919/FPL.2017.8056829
关键词:
摘要: The benefits of customising the precision throughout an FPGA design according to a tolerance are well known. However, at runtime has potential for even greater performance impact. In this paper, we add ability dynamically choose internal datapath. This enables result that is least as accurate worst-case under standard precisions, whilst internally operating lower precision. We demonstrate technique on fused floating-point dot-product circuits. show circuits with inputs have wide dynamic range, can see substantial resource savings. provide examples savings up 75% DSPs and 16% ALMs over optimised design.
sci-hub.se 参考文章(16)
Antonio Roldao Lopes, George A. Constantinides, A fused hybrid floating-point and fixed-point dot-product for FPGAs applied reconfigurable computing. ,vol. 5992, pp. 157- 168 ,(2010) , 10.1007/978-3-642-12133-3_16
Jeremy Fowers, Eric S. Chung, Karin Strauss, Greg Stitt, Kalin Ovtcharov, A High Memory Bandwidth FPGA Accelerator for Sparse Matrix-Vector Multiplication field-programmable custom computing machines. pp. 36- 43 ,(2014) , 10.1109/.21
David Boland, George Constantinides, Revisiting the reduction circuit: A case study for simultaneous architecture and precision optimisation field-programmable technology. pp. 410- 413 ,(2013) , 10.1109/FPT.2013.6718401
David Boland, George A. Constantinides, A Scalable Precision Analysis Framework IEEE Transactions on Multimedia. ,vol. 15, pp. 242- 256 ,(2013) , 10.1109/TMM.2012.2231666
Florent de Dinechin, Bogdan Pasca, Large multipliers with fewer DSP blocks field-programmable logic and applications. pp. 250- 255 ,(2009) , 10.1109/FPL.2009.5272296
Antonio Roldao, George A. Constantinides, A High Throughput FPGA-Based Floating Point Conjugate Gradient Implementation for Dense Matrices ACM Transactions on Reconfigurable Technology and Systems. ,vol. 3, pp. 1- 19 ,(2010) , 10.1145/1661438.1661439
David Boland, George A. Constantinides, Optimizing memory bandwidth use and performance for matrix-vector multiplication in iterative methods ACM Transactions on Reconfigurable Technology and Systems. ,vol. 4, pp. 1- 14 ,(2011) , 10.1145/2000832.2000834
Hani H. Saleh, Earl E. Swartzlander, A floating-point fused dot-product unit international conference on computer design. pp. 427- 431 ,(2008) , 10.1109/ICCD.2008.4751896
David B. Thomas, Wayne Luk, Estimation of sample mean and variance for Monte-Carlo simulations field-programmable technology. pp. 89- 96 ,(2008) , 10.1109/FPT.2008.4762370
Martin Langhammer, Floating point datapath synthesis for FPGAs 2008 International Conference on Field Programmable Logic and Applications. pp. 355- 360 ,(2008) , 10.1109/FPL.2008.4629963