Not that you would normally buy a cheap NVIDIA GeForce graphics card for deep learning tasks like the recently launched GTX 1050 series, as part of running some other fresh CUDA+OpenCL benchmarks I realized I hadn't run any Caffe benchmarks in a while so here are some fresh numbers today. With thirteen NVIDIA GeForce graphics cards including all the consumer GeForce GTX 1000 Pascal cards to date, here are some Caffe benchmarks using the latest NVIDIA 375.10 Linux driver on Ubuntu along with CUDA 8.0 and cuDNN.
For your viewing pleasure this Sunday are Caffe benchmarks of the GTX 680, GTX 760, GTX 780 Ti, GTX 950, GTX 960, GTX 970, GTX 980, GTX 980 Ti, GTX 1050, GTX 1050 Ti, GTX 1060, GTX 1070, and GTX 1080. Unfortunately I don't have access to any of NVIDIA's Quadro/Tesla cards for benchmarking so it's just these consumer cards to have fun with. Today are just the Caffe results while I have a lot of other CUDA 8.0 and OpenCL benchmark results to share in the next few days.
With the Caffe test profile via the Phoronix Test Suite and OpenBenchmarking.org I ran the GPU AlexNet and Googlenet tests built against CUDA 8.0 and cuDNN. For adding more value to these results are also some power consumption numbers (measured via a WattsUp Pro measuring the AC power and interfacing with the Phoronix Test Suite) along with the GPU card temperatures.