COLUMNUS - An SIMD architecture for pattern recognition and simulations of statistical physics
Martin Neschen · 2002
Many interesting problems including simulations of statistical physics, pattern recognition and neural networks can be treated efficiently by performing calculations in parallel on a large number of discrete, often even binary variables. As general-purpose computers are not well adapted to these problems, the authors have developed an SIMD array of bit-sequential processors providing an extended set of Boolean operations including an efficient bit counting. Each processor is directly connected to a DRAM memory which allows simulations of very large systems. A CMOS chip integrating 32 bit-sequential processors has been designed in 1.5 /spl mu/m. A dedicated hardware accelerates binary matrix multiplications, which are important for pattern recognition and neural network evaluations. The authors present both the hardware and interesting applications including the recognition of handwritten characters.>