Mixed-Precision Architecture for GPU Tensor Cores

NVIDIA introduced a new generation of graphics processing unit (GPU), called Volta, to meet the growing demand for higher performance in Deep Neural Networks (DNNs). Volta GPUs exploit a dedicated hardware unit, called Tensor Core (TC), to accelerate multiplication of matrices. While TCs offer significant computational horsepower for deep learning applications, they increase the power budget of DNNs. In this work, we exploit error resiliency property of DNNs and propose a technique to reduce energy of TCs. In particular, we propose an approximate architecture with the flexibility of switching between exact and approximate operating modes. In approximate mode, TCs offer significant energy saving at the cost of lower accuracy. To mitigate the impact of approximation on accuracy, we propose a mixed-precision architecture where a combination of exact and approximate units is used to reduce error in DNNs. The mixed-precision architecture provides the flexibility for the software stack to tune the level of approximation to satisfy a desired inference accuracy in DNNs. We evaluate the proposed architecture using DNNs selected from a wide range of application domains and show that the mixed-precision architecture achieves 40% energy saving while maintaining accuracy of DNNs.