Question about 4x-16x Anisotropic Texture Filtering and Performance

For example FEAR1 series has broken AF and uses bilinear AF.. to fix it you need to use ingame trilinear and force driver AF so it uses trilinear AF.


There were quite a few u3e games that had similar glitch, but you had to fiddle with cfg.ini file.. Masseffect, bulletstorm,bioshock, killing floor2 and some more.

Ingame.. driver af overrides that and uses 16x on all.



Speaking of. I force only per game profile.. there are still some apps games that don't work nice if forced globally.

 

Just leaving an update for this that I found interesting. In Red Dead Redemption 2 with Hardware Unboxed's tests for that game, increasing anisotropic texture filtering did have a measurable decrease in performance.

All things considered it was pretty minor (just a few percentage points moving from say 4x to 16x), but I'd argue the increase in visual quality from 8x to 16x is also pretty minor as well so I could see some users who want "every frame" in Overwatch or something leaving it at 4x for sake of example.

TweakTown in their overwatch tests also found it had a very small, but still measurable decrease in FPS too interestingly enough.

So, basically it seems like splitting hairs -- running 16x anisotropic filtering will result in slightly lower performance in some cases at least, but usually the impact is relatively small/only 1-2% even when going from something like 4x to 16x on modern gpus if that -- Red Dead 2 in those HWU tests seemed like the most noticeable decrease and even there it was like 2% between 4x and 16x iirc.

 

For best iq force af to 16x and clamp the negative lod bias.

 

This.
Regarding performance, I think that 16x AF is worth even a few percent; 4x and lower are especially noticeable.
Also, Horizon Zero Dawn had an AF slider that didn't work on launch, though this was patched and it works as it should now.

 

Same. Wasn't the case in the old days though

 

Very true.

 

I still force it on few games, thought at this point forceing it global should have no hit and should have no issues. some games have still have issue with AF and some dont even support it in there own settings which is odd.

 

Before 2005? Maybe. After that, performance has been negligible. I've had the driver set to force 16xAF since at least 2008 and have only ever had problems once or twice (an earlier version of GOG's client would glitch out with forced AF I can recall).
And most games that have some kind of issue, have a separate block placed on their profile by Nvidia. So you can safely force it globally and never have a problem. (Tons of games don't even have options for Texture filtering, so if you don't force it you are stuck with awful filtering)
It's absolutely and utterly absurd at this point that consoles like PS5 still don't have guaranteed 16xAF. It was absurd when PS4/XBO launched, it was absurd when PS4P/XBO-X launched and it's still just as crazy that somehow this causes significant performance issues on consoles with all their "To the metal optimization" on GPUs nearly on par with high end RTX 2000 series cards.

 

I never knew Anisotropic filtering could impact VRAM usage but Googling around I saw several articles including an old one from Gamers Nexus claiming high texture AF can become VRAM intensive since, if I’ve understood this video correctly, it’s basically storing extra pieces of the texture in memory.

The question I have then is how VRAM intensive could AF actually be? I suppose I’ve never really tested to know so I will try to load up some games and compare the VRAM usage with 1x vs 16x.

EDIT: https://www.gamersnexus.net/dictionary/7-game-graphics-settings/44-anisotropic-filtering#:~:text=Anisotropic filtering will take multiple,process that becomes VRAM-intensive.
> “Anisotropic filtering will take multiple samples per texel to determine the best appearance for the current situation. An anisotropic filtering setting of 16x (“16-tap”) means the GPU will collect 16 samples per texel, a process that becomes VRAM-intensive.”

 

Never really saw vram increase from say 4x or 16x in games so that news to me, i know performance difference between it negligible to point there no reason not to use 16x

 

It's not that intensive. I think I have been almost universally turning it on since the original Radeon/GeForce 2 era.

 

Although related. One thing are Mip Maps. Another thing is Anisotropic Filtering.
In the case of AF, I never noticed it using more vram, not in a meaningful way.
What it does use a lot, is memory bandwidth, because it requires to sample textures more times. This is from where it can drop some performance, but it only happens if the GPU is rendering a scene that already requires a lot of memory bandwidth.

 

Good to know thanks, I’ll do some testing later in Hunt Showdown or something, but yeah I can’t say I’ve noticed a big increase historically and I didn’t even know it had any impact before now.

 

Er, apologies, but what exactly is the difference between standard GPU performance and specifically memory bandwidth? I would’ve thought by the sound of it that memory bandwidth and GPU memory (VRAM) were related, are they not? I’ll do some Googling around after work. Thanks for the info that’s helpful to know.

Side note it is an irritated me when games don’t split off texture filtering level from texture resolution. Apparently CSGO had them separate while CSGO2 now is them combined so anything other than ultra uses like 4x or less going off that video I attached. Even I’m using Low textures on an old PC I’d still probably use like 4x AF since it’s generally so cheap these days.

 

It probably depends on the system i suppose. I always run it at 16x but with a 3090 i don't see any performance drops unless it's too small for me to notice.

When i had a weaker PC i would turn it down just to squeeze out as many frames as possible.

 

A GPU can have many theoretical performance metrics. Each will affect real world performance, in different situations. But they both refer to the amount of time the GPU has to wait for data.
The amount of vram is only an issue when the GPU has to get data from main memory, or SSD. As this is much slower both in latency and bandwidth, than going to vram.
The memory bandwidth the vram has, is another issue. This refers to the amount of data a GPU can fetch from vram, in a time period. Usually measured in GB/s.

Having a bottleneck in the vram bandwidth, usually only results in lower frame rate, as the GPU has to wait a few extra ms, for data.
Having to go to system memory to fetch data is much worse. Even worse it that data is not in memory, but on the SSD. In this case, the latency is much higher, and will result in a stutter.

 

Personally I have a hard time telling the difference in most scenes between 8x and 16x. 4x is alright too but you can definitely notice the compromises.

 

Thanks a lot that's helpful to know

 

I been forcing 16x in CP for as long as i can remember, I can notice huge difference in IQ with it on and off special in game that dont even use it, applying in cp is make those game better, or in case of game that have it but it broke. Could swore Legendary dragon? verison of skyrim had AF setting in cp but never worked cause I remember setting to 16x and seeing no difference and then applying it via NVCP and seeing huge diffrence, which why the profile has had that set to 16x. the current verisons of game dont have any option in games config anymore, i assume it was removed cause never worked.






I dont see much of difference when going from say 8x to 16x, hard to spot less it still pictures. but 4x to 8x the difference is obvious to me and i see virtual no difference in performance hit regardless of settings.

 

8x vs 16x can almost be game dependant as to if your going to notice at all. If it's all close quarters, and never really opens up, it may not even have a situation where you can really tell without looking for it. Where as if it's say a racing, game... those nice long roads, straight lines, extreme angles etc... your going to notice before you even leave the start line.