Achieving Native GPU Performance for Out-of-Card Large Dense Matrix Multiplication

In this paper, we illustrate the possibility of developing strategies to carry out matrix computations on heterogeneous platforms which achieve native GPU performance on very large data sizes up to the capacity of the CPU memory. More specifically, we present a dense matrix multiplication strategy on a heterogeneous platform, specifically tailored for the case when the input is too large to fit on the device memory, which achieves near peak GPU performance. Our strategy involves the development of CUDA stream based software pipelines that effectively overlap PCIe data transfers with kernel executions. As a result, we are able to achieve over 1 and 2 TFLOPS performance on a single node using 1 and 2 GPUs respectively.