Better IQ: Downscaling (AMD+NV)

[boxleitnerb]

I’ve been testing it extensively in a range of modern titles and I’ve haven’t seen any major issues with it, unlike MLAA which is almost universally unusable.

Sure, it’s not as good as 2xTrSS (especially for alpha textures) but it still provides a big improvement over no AA. It’s also extremely fast, even at very high resolutions.

IQ is very good considering it’s a shader post-filter, and it’s effective at getting AA into situations where none is possible due to compatibility and/or performance reasons.


I’ve been using OGSS modes since about 2004, so I’m very familiar with how they look. The fact is you need a minimum of 4x to provide usable increases to IQ, and you’ll drop to as low as 25% of your original performance. And even with such extreme punishment, your worst cases still only get 2xAA.

Like I said above, it’s too slow for new games and has inferior IQ compared to the combined modes, which work fine in older games. So it doesn’t really fit anywhere as having a practical use.

If you have a fast graphics card or even multiple cards, what is the harm in running OGSS via downscaling and FXAA? Together the result is quite appealing I think. Just take Max Payne as an example. MSAA is bad and very very costly. 3840x2160+FXAA gives similar performance and much better IQ. For older games I would not use it at all. There, SGSSAA is the best choice. Only in titles where no AA is possible does it really make sense.

Nowadays cards are so fast and "AA" methods so cheap (FXAA etc.), that it's often not a problem to go to 4K resolutions.

 

[Anarchist420]

No, why would they? LCDs are fixed to 60/120Hz maximum and changing the resolution makes no difference to this upper limit.

Thanks for answering my question To answer yours... I was thinking that it was possible because if it has the bandwidth required for 2560x1600 @ 60 Hz, then it would have the bandwidth required for half resolution @ 120 Hz. I wasn't aware they were absolutely fixed, although the TMDS receivers are digital so it makes sense that they'd be fixed.

 

[KingFatty]

Any performance numbers, to see what FPS you get using this method, vs. some of the other techniques in the various responses? Does this technique hurt performance on par with the quality, or is the quality too low to justify the performance hit (compared with traditional AA techniques etc.)?

 

[boxleitnerb]

Any performance numbers, to see what FPS you get using this method, vs. some of the other techniques in the various responses? Does this technique hurt performance on par with the quality, or is the quality too low to justify the performance hit (compared with traditional AA techniques etc.)?

This method doesn't make sense when traditional AA works
So no need comparing these.

I just did a quick comparison in Just Cause 2 at my current save location (again, here MSAA and SGSSAA work, this is just an example). In 1920x1080 no AA I have 90fps, in 3840x2160 no AA it is 40fps. So I would say it costs a good 50%. But remember: In games that don't support MSAA, there is additional performance headroom. Even more so if you have a highend card or SLI/CF.

 

[KingFatty]

Interesting that you quadrupled the number of pixels so you think the load on the card would increase by 400% and make the FPS drop to 25%, but the FPS only dropped from 90 to 40 (much smaller drop).

 

[boxleitnerb]

I think this is quite normal. Look up some Eyefinity or NV Surround benchmarks. In most of the cases triple resolution doesn't give you 33% of the fps, but a healthy bit more. But this is different from game to game. In Diablo 3 with Ambient Occlusion at high quality it goes from 100 to 31, so quite a difference. But 2880x1620 is perfectly playable.

 

[PowerK]

Honestly I still have not completely understood where the downscaling takes place:
In the GPU or in the monitor.

I'm inclined to say that it is a GPU/driver thing because the GT520 can do things that my GTX580 cannot, namely 3840x2160@60Hz - on the same monitor with the same cable, that is. It may have to do with the way the display driver addresses the monitor.

Good topic. This reminds of my good'ol days of using custom resolutions with SONY GDM-FW900 CRT + PowerStrip.

That being said, if the scaling is 100% done on the GPU/driver side, why do people with different monitors have different results?
I haven't had any luck with 4k resolution @60Hz with my U3011 (2560x1600 native).

 

[boxleitnerb]

Probably because it is still the GPU that has control over addressing the monitor.

Btw with Kepler things changed in terms of total horizontal pixels. With Fermi these had to be increased, with Kepler they actually should be lowered. The following settings worked for many people at 60Hz:

 

[PowerK]

That's for 1080p monitors. I tried that.. except I had my active pixels at 2560 (horizontal) and 1600 (vertical). No dice.

3840x2400p @ 30Hz worked with 'automatic' timings. But I want 60Hz.

I've also googled for downsampling experience with U3011, nothing came up, unfortunately.
I think most of U3011 users (if not all), don't bother with downsampling since 2560x1600p native resolution is already quite high.

 

[boxleitnerb]

My bad, I skipped the U3011 part.

There is a downsampling tool out there (names "SSAA tool") which works with all cards in DX10+ only. It's a bit buggy though, you have to add the executable path for each application to a text file and you need a signed driver. Often mouse coordinates are not translated correctly or the tool doesn't work at all, but it's worth a try I guess.

I'll see if I can find and upload it.