- Messages
- 572
- Country
From Cameni: (Outerra dev)
See Thread: http://forum.outerra.com/index.php?topic=3494.0
Hi everybody
A while ago we decided to make a few tests measuring the rendering performance of modern GPUs, that should help us decide the optimal rendering strategy and data layout for various types of rendering tasks an upcoming features. We ran the tests on a couple of graphics cards, and the preliminary results of these tests are very interesting, for example here are graphs from our procedural grass test:
Raw performance Mtris/s
Performance/price, Mtris/s per dollar
In order to gather more complete data, we would like to ask you to download the performance test, and after unpacking run the two test suites there: grass_test.bat and buildings_test.bat
When each tests completes, you will see your score, and you will be asked to send the results to us so we can incorporate results from different hardware into our results.
Please note that the test program requires OpenGL 4.x drivers and it may not run on some older cards that only support OpenGL 3.x. Also, the test program requires 64-bit Windows.
If you run dctest64.exe alone, you will see a different test that will be ready later. This initial build is mainly testing procedural rendering, whereas the following one will deal with model rendering performance.
You can download the test here: perftest.zip
Run grass_test.bat and after it finishes also buildings_test.bat
UPDATE: from Post 1
Yet another performance benchmark, this time focusing on finding the optimal settings for model rendering (non-procedural) across different hardware. This one is testing the achievable triangle throughput and number of draw calls in several rendering modes and with varying mesh sizes.
There are 103 tests there, each taking roughly 5 seconds, so it can take 9-10 minutes total.
If you would like to help us gathering performance data, the test can be downloaded here: outerra_perf_tests.zip
Run the cubes_test.bat
Here are some results from our cards
* Nvidia cards are able to reach much higher draw call counts than AMD (40M vs 13M), apparently having a lower overhead
* newer AMD cards need larger meshes to achieve better performance (triangle throughput), 208-triangle mesh still doesn't start to saturate it
* interesting to compare the number of draw calls in this test (single-core OpenGL) with the reported number of draw calls inDX12 (13.4 Million on multiple cores)
* indirect draw is the fastest method, for some reason faster than instanced draw call that does the same thing without an extra buffer
AMD R9 380
AMD 7970 (280X)
NV 750TI
NV 960GTX
NV Titan Z
Thanks for your help! Complete results will be published on Outerra blog together with comments.
NOTE: PRELIMINARY RESULTS FROM THE GRASS TEST ARE NOW IN THE BLOG
See Thread: http://forum.outerra.com/index.php?topic=3494.0
Hi everybody
A while ago we decided to make a few tests measuring the rendering performance of modern GPUs, that should help us decide the optimal rendering strategy and data layout for various types of rendering tasks an upcoming features. We ran the tests on a couple of graphics cards, and the preliminary results of these tests are very interesting, for example here are graphs from our procedural grass test:
Raw performance Mtris/s
Performance/price, Mtris/s per dollar
In order to gather more complete data, we would like to ask you to download the performance test, and after unpacking run the two test suites there: grass_test.bat and buildings_test.bat
When each tests completes, you will see your score, and you will be asked to send the results to us so we can incorporate results from different hardware into our results.
Please note that the test program requires OpenGL 4.x drivers and it may not run on some older cards that only support OpenGL 3.x. Also, the test program requires 64-bit Windows.
If you run dctest64.exe alone, you will see a different test that will be ready later. This initial build is mainly testing procedural rendering, whereas the following one will deal with model rendering performance.
You can download the test here: perftest.zip
Run grass_test.bat and after it finishes also buildings_test.bat
UPDATE: from Post 1
Yet another performance benchmark, this time focusing on finding the optimal settings for model rendering (non-procedural) across different hardware. This one is testing the achievable triangle throughput and number of draw calls in several rendering modes and with varying mesh sizes.
There are 103 tests there, each taking roughly 5 seconds, so it can take 9-10 minutes total.
If you would like to help us gathering performance data, the test can be downloaded here: outerra_perf_tests.zip
Run the cubes_test.bat
Here are some results from our cards
* Nvidia cards are able to reach much higher draw call counts than AMD (40M vs 13M), apparently having a lower overhead
* newer AMD cards need larger meshes to achieve better performance (triangle throughput), 208-triangle mesh still doesn't start to saturate it
* interesting to compare the number of draw calls in this test (single-core OpenGL) with the reported number of draw calls inDX12 (13.4 Million on multiple cores)
* indirect draw is the fastest method, for some reason faster than instanced draw call that does the same thing without an extra buffer
AMD R9 380
AMD 7970 (280X)
NV 750TI
NV 960GTX
NV Titan Z
Thanks for your help! Complete results will be published on Outerra blog together with comments.
NOTE: PRELIMINARY RESULTS FROM THE GRASS TEST ARE NOW IN THE BLOG
Last edited: