NVIDIA GeForce vs. AMD Radeon Vulkan Neural Network Performance With NCNN

With having added Tencent's NCNN tests to the Phoronix Test Suite with Vulkan acceleration, here is a look at the real-world impact by using RealSR-NCNN for scaling up with RealSR. Various NVIDIA GeForce and AMD Radeon graphics cards were tested for this initial NCNN / RealSR-NCNN Vulkan comparison.

This is our first time looking at how well Vulkan performs in this area with the current state of the Linux drivers. The GeForce hardware was tested with the latest 450 series proprietary driver while on the Radeon side it was with Linux 5.9 and Mesa 20.3-devel using the RADV Vulkan driver. One of the Tencent developers working on NCNN has commented as well that using RADV's ACO offers a big boost for the performance, which fortunately is the current default for the RADV Vulkan driver.

The test image...

RealSR-NCNN was run on all of the GPUs as well for its image upscaling to 4x the original size. For those curious here is the original image and then the upscaled version, accelerated with Vulkan.

Based on the recent, available graphics cards I have for testing the ones listed below were tested. Alas no RTX 30 series coverage yet as still waiting on hardware from NVIDIA. But additional NCNN/RealSR-NCNN tests to follow.

- GTX 980 Ti
- GTX 1060
- GTX 1070
- GTX 1080
- GTX 1650
- GTX 1650 SUPER
- GTX 1660
- GTX 1660 SUPER
- GTX 1660 Ti
- RTX 2060
- RTX 2060 SUPER
- RTX 2070
- RTX 2070 SUPER
- RTX 2080
- RTX 2080 SUPER
- RTX 2080 Ti
- TITAN RTX
- RX Vega 56
- RX 5600 XT
- RX 5700
- RX 5700 XT
- Radeon VII

That's the gist of this initial RealSR-NCNN Vulkan testing with other NCNN tests forthcoming via the Phoronix Test Suite.