MAIN

Vulkan is not always faster than OpenGL!

Upon my research and studying over Vulkan reaching the point of Rendering and Presentation chapter of vulkan-tutorial.com. I took a chance to compare the performance of rendering a single triangle between Vulkan, and OpenGL. Both have at most similar setup as much as possible. You can see my initial comment over this topic here.

Surprisingly, I got lower FPS gained from Vulkan. It is lower by ~2000 FPS as compared to OpenGL. To be clear, Vulkan gets ~7800 FPS, and OpenGL gets ~9800 FPS. You can check my OpenGL code for such setup here.

Although I slightly modified the synchronization approach used based on top of such chapter to follow along with approach used in GL_vs_VK - Test1, in my opinion it is easier to understand and see why it would need such fences, or semaphores. With that, it helps boosting up FPS by ~200 FPS. It still doesn’t able to match with OpenGL’s result.

GL_vs_VK is a thesis project by RippeR37 comparing performance between OpenGL and Vulkan graphics API under different circumstance both for single-threaded and multi-threaded.

So I further proceeded with GL_vs_VK to see whether the results align with what I’ve found if I would modify the number of triangle to render.

Here is the results I got. Notice at number of triangle is 1, Vulkan is still at worse than OpenGL. Reasons for this will come after the listing of the benchmark results.

Setup

My machine is Ubuntu 19.10, kernel: 5.3.18-050318-generic, with Ryzen 3 2200g with 16 GB RAM, no discrete GPU.

By making a modification to numbers of triangle in src/tests/test1/BaseBallsSceneTest.cpp to be desire number, then execute 4 sets of benchmark as follows.

OpenGL Single-Threaded

./GL_vs_VK -t 1 -api gl -benchmark -time 10

OpenGL Multi-Threaded

./GL_vs_VK -t 1 -api gl -benchmark -time 10 -m

Vulkan Single-Threaded

./GL_vs_VK -t 1 -api vk -benchmark -time 10 -m

Vulkan Multi-Threaded

./GL_vs_VK -t 1 -api vk -benchmark -time 10 -m

Notice that we keep it running for 10 seconds.

Benchmarks

GL_vs_VK - Triangle Count 1

OpenGL

Threading   Min ft (ms)   Max ft (ms)   Avg ft (ms)   Max FPS   Min FPS   Avg FPS
Single-Threaded 0.061760 11.682720 0.168683 16191.709844 85.596505 5928.292623
Multi-Threaded 157.958540 195.491431 173.885831 6.330775 5.115314 5.750900

Vulkan

Threading   Min ft (ms)   Max ft (ms)   Avg ft (ms)   Max FPS   Min FPS   Avg FPS
Single-Threaded 0.049210 2.674723 0.266203 20321.072953 373.870491 3756.527258
Multi-Threaded 0.162930 8.399509 0.295276 6137.605106 119.054578 3386.665932

Note: Red text indicates when OpenGL has better performance than Vulkan.

GL_vs_VK - Triangle Count 100

OpenGL

Threading   Min ft (ms)   Max ft (ms)   Avg ft (ms)   Max FPS   Min FPS   Avg FPS
Single-Threaded 0.187840 5.085851 0.288568 5323.679727 196.623928 3465.385333
Multi-Threaded 0.273139 7.251312 0.427964 3661.139566 137.906078 2336.646718

Vulkan

Threading   Min ft (ms)   Max ft (ms)   Avg ft (ms)   Max FPS   Min FPS   Avg FPS
Single-Threaded 0.097439 6.650845 0.268783 10262.831105 150.356834 3720.467167
Multi-Threaded 0.182610 5.852518 0.325074 5476.151361 170.866625 3076.224899

Note: Avg FPS of both graphics APIs indicate the number of triangles around ~1xx is the diminishing return point at which Vulkan will start to shine.

GL_vs_VK - Triangle Count 200

OpenGL

Threading   Min ft (ms)   Max ft (ms)   Avg ft (ms)   Max FPS   Min FPS   Avg FPS
Single-Threaded 0.282030 2.357284 0.391126 3545.722086 424.217023 2556.718589
Multi-Threaded 0.327139 6.800143 0.526437 3056.804600 147.055731 1899.562859

Vulkan

Threading   Min ft (ms)   Max ft (ms)   Avg ft (ms)   Max FPS   Min FPS   Avg FPS
Single-Threaded 0.148400 5.213387 0.288929 6738.544474 191.813882 3461.061541
Multi-Threaded 0.201749 3.371651 0.354996 4956.654060 296.590602 2816.929524

GL_vs_VK - Triangle Count 500

OpenGL

Threading   Min ft (ms)   Max ft (ms)   Avg ft (ms)   Max FPS   Min FPS   Avg FPS
Single-Threaded 0.602639 6.663223 0.736657 1659.368212 150.077523 1357.482968
Multi-Threaded 0.582768 5.724255 0.920224 1715.948714 174.695222 1086.692188

Vulkan

Threading   Min ft (ms)   Max ft (ms)   Avg ft (ms)   Max FPS   Min FPS   Avg FPS
Single-Threaded 0.318399 6.668269 0.363154 3140.713382 149.963956 2753.650049
Multi-Threaded 0.243349 4.082437 0.397493 4109.324468 244.951729 2515.765958

GL_vs_VK - Triangle Count 1000

OpenGL

Threading   Min ft (ms)   Max ft (ms)   Avg ft (ms)   Max FPS   Min FPS   Avg FPS
Single-Threaded 1.121107 11.794929 1.343919 891.975521 84.782198 744.092734
Multi-Threaded 0.993487 6.283654 1.555241 1006.555697 159.143072 642.98

Vulkan

Threading   Min ft (ms)   Max ft (ms)   Avg ft (ms)   Max FPS   Min FPS   Avg FPS
Single-Threaded 0.598028 5.253233 0.650044 1672.162507 190.358966 1538.357590
Multi-Threaded 0.313759 5.227853 0.476048 3187.159572 191.283114 2100.627088

Summary

Imagine 3D-based virtualization application needs to render numbers of mesh on screen thus it needs to send vertex attribute as data to GPU each frame if those data are dynamic. If such application isn’t reaching CPU-bound point (the diminishing return point of benefit from OpenGL) then Vulkan-based application doesn’t shine just yet and will perform not quite fully to its low overhead abstraction.

So be aware of such CPU-bound point of the application. OpenGL and Vulkan will offer different benefit and shine at different scenario.

Aside, multi-threading in Vulkan will shine and show its benefit when the number of instances to draw on screen is high enough to compensate with some costs used in handling or setting up multi-threading ecosystem. Notice at triangle count of 1000.



First published on April, 13, 2020






Written by Wasin Thonkaew
In case of reprinting, comments, suggestions
or to do anything with the article in which you are unsure of, please
write e-mail to wasin[add]wasin[dot]io

Copyright © 2019 Wasin Thonkaew. All Rights Reserved.