 |
Center for Nonlinear Studies |
Randomness is valuable for numerous applications in computing. Hardware implementations of a wide range of applications in areas such as stochastic neural networks and machine vision can be realized with much lower circuit complexity if randomness is leveraged. Applications that rely on the simulation of stochastic phenomena, such as computational biology and quantum physics, also exploit randomness. In this talk, a novel paradigm for approximate computation called Stochastic Computing (SC) is advocated and its issues will be addressed to open up novel applications and paradigms in randomized computation. The goal of this research is to develop a novel SC framework named clQSC (OpenCL Quasi-Stochastic Computing) and justify its performance, efficiency and viability in emerging application domains such as stochastic neural networks, machine vision and deep learning. clQSC is an innovative approach fusing: 1) massive fine-grained parallelism and scalability of FPGA; 2) programmability of OpenCL; and 3) excellent progressive precision of LD (Low Discrepancy) sequences; into a synergistic and holistic stochastic computing framework to open up new application domains which have never been possible in conventional SC. The proposed clQSC framework can be used to quickly and automatically design, optimize and synthesize custom parallel and scalable SC hardware fabric with efficient LD (Low Discrepancy) sequence generators from C-like OpenCL kernel codes for extremely high efficiency, throughput and progressive precision - without even knowing HDL (Hardware Description Language) such as VHDL and Verilog. Also, flexibility and configurability of the parallelization paradigm of the proposed clQSC will trailblaze unique SC applications which have never been possible.
Host: Chris Rawlings