Earlier this week I posted some benchmarks of GPUOpen's new Radeon Open Compute ROCm OpenCL stack that premiered last month and they are working to make completely open-source. In those initial benchmarks I compared the ROCm 1.4 OpenCL performance to the existing AMDGPU-PRO OpenCL implementation on Linux. For those wondering how these two Radeon OpenCL stacks compare to NVIDIA, here are some fresh benchmarks.
After finishing those OpenCL benchmarks for the earlier article I proceeded to pop in the line-up of my available GeForce GTX 1000 "Pascal" graphics cards and benchmark them using the NVIDIA 375.26 driver. The same Ubuntu 16.04 LTS x86_64 system was used for testing with Xeon E3-1280 v5 Skylake CPU.
The NVIDIA tests were with the GTX 1050, GTX 1050 Ti, GTX 1060, GTX 1070, and GTX 1080. The AMDGPU-PRO / ROCm tests were with the Radeon RX 460, RX 480, and R9 Fury. The ROCm OpenCL implementation is currently only compatible with Polaris and Fiji graphics processors.
This is just a comparison of the NVIDIA OpenCL (v1.2) stack compared to AMD's ROCm 1.4 / AMDGPU-PRO drivers and not any CUDA vs. ROCm comparison or anything along those lines. The same benchmarks were used of SHOC, Rodinia, Darktable, JuliaGPU, MandelGPU, MandelbulbGPU, and LuxMark. All of these OpenCL benchmarks were facilitated in a fully-automated and reproducible manner using the open-source Phoronix Test Suite benchmarking software.