ReFloat: Low-Cost Floating-Point Processing in ReRAM

We propose ReFloat, a principled approach for low-cost floating-point processing in ReRAM. The exponent offsets based on a base are stored by a flexible and fine-grained floating-point number representation. The key motivation is that, while the number of exponent bits must be reduced due to the exponential relation to the computation latency and hardware cost, the convergence still requires sufficient accuracy for exponents. Our design reconciles the conflicting goals by storing the exponent offsets from a common base among matrix values in a block, which is the granularity of computation in ReRAM. Due to the value locality, the differences among the exponents in a block are small, thus the offsets require much less number of bits to represent exponents. In essence, ReFloat enables the principled local fine-tuning of floating-point representation. Based on the idea, we define a flexible ReFloat format that specifies matrix block size, and the number of bits for exponent and fraction. To determine the base for each block, we propose an optimization method that minimizes the difference between the exponents of the original matrix block and the converted block. We develop the conversion scheme from default double-precision floating-point format to ReFloat format, the computation procedure, and the low-cost floating-point processing architecture in ReRAM.