An evaluation of contemporary heterogeneous computing platforms for data intensive applications

The end of Dennard scaling is forcing innovation in computing architectures; the old regime, where exponential benefits accrued by using a newer process technology, no longer holds. One path of innovation is to exploit heterogeneous platforms, and match applications to the platforms on which they execute efficiently. Heterogeneous processing in the server domain is a big hardware/software design space. As a result, research that looks narrowly at one or two applications, and hardware options does not usually generalize. In this work, we present a broad comparison of a variety of hardware accelerators and applications. We investigate an Intel Xeon E5-2620, ARM Cortex A15, Nvidia Tegra Jetson K1, Nvidia Titan X, Intel Xeon Phi coprocessor, and Xilinx Zynq FPGA, and evaluate how efficiently they execute key data mining applications. We focus on energy efficiency, but also reflect on the developer effort required to utilize each platform. At the highest level, the results mirror conventional wisdom: FPGAs are about as efficient as GPUs which are more efficient than low-power processors. Developer effort is the exact opposite; processors are the easiest to program, followed by GPUs and then FPGAs. If the producers of more energy-efficient compute platforms (FPGAs, GPUs) wish to be ubiquitous in the datacenter, they (still) must work on improving their programming environments.