Float-Fix: An Efficient and Hardware-Friendly Data Type for Deep Neural Network

Recent years, as deep learning rose in prominence, neural network accelerators boomed. The existing research shows that both speed and energy-efficiency can be improved by low precision data structure. However, decreasing the precision of data might compromise the usefulness and accuracy of the underlying AI. And the existing studies can not meet all AI application requirements. In the paper, we propose a new data type, called Float-Fix (FF). We introduce the structure of FF and compare it with other data types. In our evaluation, the accuracy loss of 8-bit FF is less than 0.12% on a subset of known neural network models, 7\(\times \) better than fixed-point, DFX and floating-point on average. We implement the hardware architectures of operators and neural processing unit using 8-bit FF data type with TSMC 65 nm Gplus High VT library. The experiments show that the hardware cost of convertors converting between 16-bit fixed-point and FF is really small. And the multiplier of 8-bit FF only needs 1188 \(\upmu \mathrm{m}^2\) area, which is nearly 8-bit fixed-point. Comparing with the neural processing unit of DianNao, FF reduces 34.3% area.