NVIDIA vs. AMD Linux Performance For GRID Autosport

[philips]

Windows vs SteamOS, AMD vs Nvidia @ 4K:

[Tomin]

Not exactly AMD, but you might wan't to read ex-nvidia employee's rant about game makers and what IHVs have to do to fix broken game code:
http://www.gamedev.net/topic/666419-.../#entry5215019

That is an interesting text to read. Thank you!

[justmy2cents]

I'm a software engineer and I know too well that this world is not different from other worlds: there's no free lunch.

so true

Right now, a lot of headaches are caused by developers doing crazy stuff in their engine and than forcing nvidia and AMD to implement workarounds in their drivers and/or profiles. And now Vulkan puts more power in the hands of the developers. Whether all developers will wield that power right remains to be seen. We will probably see a lot of good things come out of Vulkan, but I'm expecting a long-ish transition period before that happens.

again, true. except you have to take it with a grain of salt that older DX and OpenGL have some inherent design problems which seem to be fixed in DX12 and Vulkan

problem you mention is much worse because of that. as similar example, just try giving those developers slightly broken compiler and then watch all the crazy workarounds they are doing. this is more like cause and effect

[bug77]

only problem with unigine is that it is not used in games much, so performance optimizations per os could really vary. add to that the fact that source is not available... so, you will get some results, but no reasoning

with UE4 (or any open source engine, except the quantity and quality of testing could really vary, where UE4 is top of the line) on the other hand any side can take open source demo like Epic provides them and then dissect what engine on certain os does not optimal if they want. this benchmark will be tractable and verifiable much more than any other

Well, I'm not saying Unigine is the go-to benchmark. Just that it's a benchmark that uses modern features and runs natively on both Linux and Windows. I'm using them (Heaven and Valley) to compare results and it shows me nvidia is just a tad slower on Linux. I'd have no issue adding something UT4-based to my testing.

and there is a matter of waiting for Vulkan which solves all the OpenGL problems

I'm a software engineer and I know too well that this world is not different from other worlds: there's no free lunch.
Right now, a lot of headaches are caused by developers doing crazy stuff in their engine and than forcing nvidia and AMD to implement workarounds in their drivers and/or profiles. And now Vulkan puts more power in the hands of the developers. Whether all developers will wield that power right remains to be seen. We will probably see a lot of good things come out of Vulkan, but I'm expecting a long-ish transition period before that happens.

[tuke81]

Does anybody know what the profiles in AMD drivers actually do? Why they speed up things so much and couldn't drivers work acceptably without them?

Not exactly AMD, but you might wan't to read ex-nvidia employee's rant about game makers and what IHVs have to do to fix broken game code:
http://www.gamedev.net/topic/666419-.../#entry5215019

[justmy2cents]

Well, we do have the Unigine Benchmarks. There's also Oil Rush, but for some reason Michael isn't using it in his benchmark suite (probably lack of CLI controls again).

only problem with unigine is that it is not used in games much, so performance optimizations per os could really vary. add to that the fact that source is not available... so, you will get some results, but no reasoning

with UE4 (or any open source engine, except the quantity and quality of testing could really vary, where UE4 is top of the line) on the other hand any side can take open source demo like Epic provides them and then dissect what engine on certain os does not optimal if they want. this benchmark will be tractable and verifiable much more than any other

and there is a matter of waiting for Vulkan which solves all the OpenGL problems

[juno]

Does anybody know what the profiles in AMD drivers actually do? Why they speed up things so much and couldn't drivers work acceptably without them?

I ask this for myself all the time. It's not only on OGL/Linux, but also for D3D games on Windows. Both, amd and nvidia, occasionally provide new driver versions for Windows when new AAA games are released.

[Tomin]

Does anybody know what the profiles in AMD drivers actually do? Why they speed up things so much and couldn't drivers work acceptably without them?

[bug77]

so true

people just expect that it will run better just because it is on linux. like any porter would ever go for full blown rewrite. with late port, at best you can hope is that they created translation layer that is strapped under the game and maybe rewrite few really expensive parts. usually port is no different than running in wine if not even worse if wine features better translation layer than the game

first games that will produce results that can be really compared will be UE4 since it features native GL without translation and has source available where tracking why and how is possible.

Well, we do have the Unigine Benchmarks. There's also Oil Rush, but for some reason Michael isn't using it in his benchmark suite (probably lack of CLI controls again).

[kernelOfTruth]

This is really underwhelming !

Yesterday it worked smoothly with around 50-60 FPS, now it was struggling with 24-40 FPS

only a reboot in between

Thought it was a kernel issue, so did re-compile several kernels and it got sometimes slightly better or worse

and that with a GTX 760 which is supposed to run at 63-64 FPS on Windows 7 or 8.1 with high+ settings, 4x MSAA and High-Res


This game suffers the same (?) fate like the new Need For Speed Hot Pursuit:

under Windows the game oftentimes needs to be set via affinity to run fixed on a certain core (CPU-bound ?!!) and priority set to high - then it runs smooth as 60+ FPS


Running GRID Autosport even with

LC_ALL=C __GL_THREADED_OPTIMIZATIONS=1 schedtool -I -e %command%

and setting nice to -5

Re-prioritizing IRQs and running via threadirqs

doesn't really help it still stutters quite suddenly during races - and that only offline for now - which really sucks,

projecting that onto Online Competitions making it not really fun


Some of the kernel and other driver devs needs to get their act together and focus on performance + latency, BOTH

Perhaps 4.4 rt-kernel will be able to help in that regard, but I doubt it

taskset (affinity for one specific core and the game) might give the highest gain if the CPU can run fully in Turbo-Mode


Hopefully Feral Interactive further tweaks and optimizes this port


edit:

Looks like it's the NVIDIA drivers this time:

https://www.reddit.com/r/linux_gamin...rid_autosport/

I'm also running 358.16 (4.3+ kernel, xorg 1.18) - so stuck there

Luckily no hardlocks yet, *knock on wood*