Yesterday, on the launch-day for the GeForce GTX 1080 Ti (GP102) graphics card, I posted GTX 1080 Ti OpenGL and Vulkan benchmarks while for those more interested in GPU compute performance, here are some preliminary OpenCL compute results.
A follow-up article will focus on CUDA performance for the GeForce GTX 1080 Ti while this article is just making use of OpenCL for GPGPU capabilities. Officially the NVIDIA Linux driver, sadly, only exposes OpenCL 1.2 but there are the beta OpenCL 2.0 extensions being exposed. Hopefully in upcoming driver updates we'll see official OpenCL 2.0+ support from the NVIDIA proprietary driver.
The GeForce GTX 1080 Ti can make for mighty powerful OpenCL/CUDA capabilities with having 3584 CUDA cores and clock speeds above 1500MHz when boosted. Graphics performance at least shows the GTX 1080 Ti being capable of beating out the GeForce GTX TITAN X (Pascal), at least under Windows, but unfortunately I don't have that card for testing. For this OpenCL comparison I used a GeForce GTX 680, GTX 780 Ti, GTX 980, GTX 980 Ti, GTX 1060, GTX 1070, GTX 1080, and GTX 1080 Ti. All cards were running at stock speeds and using the NVIDIA 378.13 driver.
No Radeon OpenCL results could be carried out, unfortunately, since their AMDGPU-PRO / ROCm driver stack doesn't yet support Ubuntu 17.04 as used for this testing (or even Ubuntu 16.10 for that matter) and their open-source Radeon OpenCL driver isn't into shape for benchmarking yet. A variety of OpenCL benchmarks from SOC to Blender and Darktable were used for benchmarking. All of the tests were streamlined and automated using the Phoronix Test Suite benchmarking software. Power consumption and performance-per-Watt metrics were also recorded by the Phoronix Test Suite.