More Vulkan NCNN Inference Benchmarks On AMD Radeon vs. NVIDIA GeForce Under Linux

Given the interest from the RealSR-NCNN Vulkan benchmarks on various NVIDIA and AMD Radeon graphics cards looking at this neural network inference framework with the task of upscaling an image by 4x the resolution using RealSR, here are some more benchmarks of the NCNN framework accelerated by Vulkan on different GPUs under Ubuntu Linux.

The numbers this weekend are of Vulkan NCNN with different AMD Radeon and NVIDIA GeForce graphics cards I had access to and running inference benchmarks on various models as opposed to the real-world RealSR-NCNN performance numbers from a few days ago.

The same system was used and the Radeon cards were with Linux 5.9 + Mesa 20.3-devel (thus RADV+ACO) while the GeForce GPUs were on the latest 450 series driver.

While the Radeon VII graphics card won the RealSR-NCNN testing and even outperformed the TITAN RTX, with this NCNN testing in some models there were regressions in performance for the Vega based graphics cards while the Navi parts actually came out ahead. But whether this is due to Navi vs. Vega hardware differences or rather RADV/ACO optimization differences is still to be tested.


For some models, NVIDIA was outright faster.




Those wanting to dig through further NCNN Vulkan benchmarks in this comparison along with GPU thermal and power stats can see this OpenBenchmarking.org result file. The NCNN test profile page has additional metrics available including for the CPU-based performance metrics for this Tencent framework.