Will GPUs with 4GB VRAM age poorly?

[sebbbi]

Yeah, NV had this to say about the matter:

That first-take on the idea rendered the same number of depth samples as a full-resolution 1920x1080 picture, but only half the shaded samples, improving performance at the expense of image quality. This manifested as a reduction in the quality and visibility of Ambient Occlusion shadowing, increased shader aliasing, decreased lighting and shading fidelity, and a loss of fidelity on smaller game elements, such as leaves, grass, visual effects and minute pieces of geometry.

That NV quote seems to be discussing a different technique, called spatial checkerboarding. Rainbow Six Siege is using temporal checkerboarding. Spatial checkerboarding is a single frame technique (no reprojection). It results in native edge sharpness, but shading (inside surfaces) is 2x lower resolution -> textures and lighting both look low res.

Temporal checkerboarding reprojects last frame content and combines that with current frame content. As long as the reprojection succeeds, this results in very close to native results (and 100% perfect result when camera is still). Temporal checkerboarding doesn't require native resolution depth buffer (or id buffer) like spatial checkerboarding, because the temporal version can combine color and depth data from two frames (alternating checkerboard pattern = 100% pixels).

I don't see any reason why SSAO should be noticeably lower quality with temporal checkerboarding. But I see lots of potential (difficult to solve) implementation reasons that might cause this issue.

 

[OlegSH]

That NV quote seems to be discussing a different technique, called spatial checkerboarding.

Nope, it's about temporal checkerboarding

In simplified terms, checkerboard rendering upscales the resolution from 1920x1080 to 3840x2160, and uses data from the previous frame to create new detailed pixels that fill in the blanks between the pixels that were upscaled, avoiding the blurriness that would otherwise occur.

I just find the term "upscaling" confusing when applied to the 2xMSAA temporal checkerboard rendering even though it's described exactly like "upscaling" almost in every press article out of here.

MSAA samples with checkerboard rendering are mapped on the output resolution pixels positions so that each 2x MSAA sample is at the centroid position of each output resolution pixel, half of pixels are obviously missing because MSAA is 2x, hence the checkerboard pattern across the screen, but press for whatever reason decided to call this "upsampling" even though there is no any upsampling involved in its traditional sense

 

[sebbbi]

MSAA samples with checkerboard rendering are mapped on the output resolution pixels positions so that each 2x MSAA sample is at the centroid position of each output resolution pixel, half of pixels are obviously missing because MSAA is 2x, hence the checkerboard pattern across the screen, but press for whatever reason decided to call this "upsampling" even though there is no any upsampling involved in its traditional sense

Yes, temporal checkerboard does no upsampling. Spatial checkerboard does. Scanline intelaced rendering isn't upsampling either (odd/even scanlines interleaved). However both of these temporal techniques also fallback to upsampling when the reprojection fails. In this case checkerboard interpolates from "+" neighborhood (both X and Y neighbors = 4 neighbors), and scanline interleaving interpolates from +-Y neighbors (= 2 neighbors). This is one of the reasons why checkerboard is a better than scanline interleaving.

 

[OlegSH]

In this case checkerboard interpolates from "+" neighborhood (both X and Y neighbors = 4 neighbors), and scanline interleaving interpolates from +-Y neighbors (= 2 neighbors). This is one of the reasons why checkerboard is a better than scanline interleaving

I used to think of it as of interpolation, traditional upsampling is much closer to extrapolation which is less accurate by definition

 

[sebbbi]

I used to think of it as of interpolation, traditional upsampling is much closer to extrapolation which is less accurate by definition

Technically they are both interpolating, since endpoints are known and the value lies between them. Checkerboard interpolation has "+" neighbors, while bilinear (center sample) has "X" neighbors. The main difference is that temporal checkerboard only needs to interpolate when reprojection fails (usually less than 10% of pixels). Other 90%+ pixels are reconstructed (not interpolated). However reconstruction usually also does some (usually higher order) interpolation (as motion vectors don't always point to exact pixel centers).

 

[DavidGraham]

Resident Evil 7 exceeds 5GB of VRAM usage, which severely castrate FuryX and the 1060 behind all the 8GB cards.

http://pclab.pl/art72889-3.html

 

[swaaye]

The 3.5 GB stricken GTX 970 looks pretty good though.

 

[seahawk]

If you look at the min. FPS the 1050ti is nearly on the Fury level. I think AMD must have some problems of their own. (Also look at 980 4GB and even the 970)

 

[DavidGraham]

HardOCP reports stuttering and hitches on the FuryX in Fallout 4 using the High Res texture pack. Apparently 4GB is not enough in this title. At least 6GB is recommended.

The only one card we felt some stutter and hitching on as we played was the AMD Radeon R9 Fury X with 4GB of dedicated VRAM on board. This simply wasn’t enough for smooth gameplay, even with its use of Dynamic RAM. Nothing beats the speed of local VRAM. We therefore would make an educated guess and suggest that other video cards with 4GB of local VRAM might also encounter this stutter or hitching loading new assets in the game as you move through it.

We would suggest a 4GB video card may not be the best option for the High Resolution Texture Pack. Based on our testing, we feel a video card with 6GB and up is suited best to run the High Resolution Texture Pack with absolute smoothness. This is not to say the game was unplayable on the Radeon R9 Fury X, it was playable, let’s make that clear, only that some may find that stutter or hitching to be annoying while playing. The Fury X worked, but it was far from "the best" experience.

http://www.hardocp.com/article/2017/02/12/fallout_4_high_resolution_texture_pack_review/13

Note that the average fps of the card was not affected, also VRAM usage topped at 3.7GB, however the FuryX was swapping a lot of textures from the SSD to VRAM and vice versa, this manifested as stutters and hitches, which never happened on the 6GB 1060, or the 8GB 480. This mirrors a behavior common in Rise of The Tomb Raider as well (especially in wide areas, and during extended gameplay sessions).

Though there technically wasn’t a performance difference in terms of framerate, we definitely felt the game stutter from time to time as it was loading in new assets. We mention this stutter on this video card because we did not experience it the three other video cards. We also only experienced it on the Fury X when we had the High Resolution Texture Pack installed. We also noticed the hard drive (in our case an SSD) was churning or loading a lot more with the Fury X installed with the High Resolution Texture Pack. We noticed there were times the hard drive light was running solid, for lengthy times, way more than any other video card we tested.

http://www.hardocp.com/article/2017/02/12/fallout_4_high_resolution_texture_pack_review/3

 

[Malo]

Resident Evil 7 exceeds 5GB of VRAM usage, which severely castrate FuryX and the 1060 behind all the 8GB cards.
http://pclab.pl/art72889-3.html

The difference between the 4gb and 8gb 480 at only 1080p is massive. A clear winner why 4gb is not enough.

My question is why this title is using so much VRAM at 1080p though? No msaa and the textures in bio hazard are very low resolution.

 

[DavidGraham]

The difference between the 4gb and 8gb 480 at only 1080p is massive. A clear winner why 4gb is not enough.

Even 6Gb is not enough with this game, the 8GB 480 is much faster than the 6GB 1060 and almost as fast as the 6GB 980Ti. The game scales well with memory size! The 3GB 1060 is completely crushed here.

why this title is using so much VRAM at 1080p though?

The game is a collection of large rooms/halls connected together through closed doors, that you frequently go through back and forth. The inside of these interiors is full of details and objects. This requires keeping all of these things close by in the VRAM for quick access, and to avoid pop ins and streaming issues. It also has to do with an option called Shadow Cache, which caches your shadows in the VRAM to avoid pop ins, and to reuse them in accelerating rendering. So that the more shadows you have cached, the more performance you will get, at least theoretically. Which increases VRAM load even more.

 

[Blazkowicz]

Thanks for describing the game.
An issue is that to see the full picture, we need the benchmark results with Shadow Cache disabled. Tanking your framerate by blitting stuff over PCIe constantly so as to "accelerate" your shadows defeats the point.

I had seen a link that gives such results, on vs off, same 1080p resolution tested.
http://www.tomshardware.fr/articles/test-resident-evil-re7-benchmark,2-2618-3.html
The boards are R9 390, GTX 1060 6GB, RX480 8GB, GTX 970, RX470 4GB, GTX 1060 3GB.

Leaving this off seems to make 4GB cards adequate and 3GB cards bearable.

On > 4GB cards, enabling shadow cache leaves average framerates about the same but minmum framerate is nicely up (btw, different benchmarking, different results...)
There is the same behavior where AMD 4GB cards are a bit weak on minimum framerate, even with the cache off.

Albeit min framerate is a weak measurement in itself, a single outlier frame will make it look bad. A decade or two ago, minimum framerate was too inconsistent or meaningless to be considered.

 

[lanek]

Resident Evil 7 exceeds 5GB of VRAM usage, which severely castrate FuryX and the 1060 behind all the 8GB cards.

http://pclab.pl/art72889-3.html

It is even worst if you enable shadow cache....

after lots of weird results with graphics card that have less then 4 GB of VRAM we found out that the problematic settings is Shadow cache. If you have an up-to 4 GB graphics card, please turn it off and you'll be playing properly. The charts are now updated to reflect that.

(result shadow cache off )

At 4K--- there the 4GB show their limit. ( need 5GB+ effectively. ).

But Resident Evil have a lot of inconstency when benchmarked.... (performance wise ).

 

[DavidGraham]

But Resident Evil have a lot of inconsistency when benchmarked.... (performance wise ).

Can't say I trust Guru3D that much, their testing has been sloppy recently and frequently, with many updates to their reviews due to test system errors. Anyhow the inconsistency steamed from them testing 4GB cards with Shadow Cache on, which caused their performance to tank. They should have tested 4GB cards with the Cache off, and 8GB cards with the Cache On.

It is even worst if you enable shadow cache....

On my 1070, Shadow Cache boosts my fps by about 8-10 frames just standing still and looking at the same scenery.

 

[lanek]

Can't say I trust Guru3D that much, their testing has been sloppy recently and frequently, with many updates to their reviews due to test system errors. Anyhow the inconsistency steamed from them testing 4GB cards with Shadow Cache on, which caused their performance to tank. They should have tested 4GB cards with the Cache off, and 8GB cards with the Cache On.


On my 1070, Shadow Cache boosts my fps by about 8-10 frames just standing still and looking at the same scenery.

Yes i know for guru3D.

Same on 480 and 390x 8GB ... But on GPU's who have only 4GB it will completely cripple the performance.

shadow cache off.
http://www.pcgameshardware.de/Resid.../Specials/Benchmark-PC-Anforderungen-1219005/

The problem is with this game, reviewers should show both result ( with cache on and off ), instead of choose one method over the other.

 

[Putas]

Albeit min framerate is a weak measurement in itself, a single outlier frame will make it look bad. A decade or two ago, minimum framerate was too inconsistent or meaningless to be considered.

Only in case of bad benchmarks practices, which can invalidate other metrics as well. I found results of 20 years old cards very consistent.

 

[Ike Turner]

Relevant:

Fallout 4 Very High Texture Pack: http://gamegpu.com/rpg/роллевые/fallout-4-high-resolution-texture-pack-test-gpu

HOCP reports stuttering that doesn't affect framerate (??) on Fiji GPUs.. but I would like to have a non biased (I wouldn't trust anything Kyle says about AMD/Radeon products..) third party report on this matter..anyone else with a Fury X or Nano can please share their experience with Fallout 4+High texture pack?

For Honor Beta: http://gamegpu.com/action-/-fps-/-tps/for-honor-beta-test-gpu

Halo Wars 2: http://gamegpu.com/rts-/-стратегии/halo-wars-2-test-gpu

...Fury X... ¯\_(ツ)_/¯

 

[Anarchist4000]

HOCP reports stuttering that doesn't affect framerate (??) on Fiji GPUs.. but I would like to have a non biased (I wouldn't trust anything Kyle says about AMD/Radeon products..) third party report on this matter..anyone else with a Fury X or Nano can please share their experience with Fallout 4+High texture pack?

Sounds like a driver bug or something. He said the Fury was using nearly 1GB (~4.8GB vs 5.8GB) less VRAM (dedicated + dynamic) than a 480 and the SSD was going crazy. So something important was getting evicted from even system memory. Otherwise I'd expect the SSD usage to be the same between cards and similarly allocated VRAM. That being the case, I doubt he was wrong about occasional stuttering.

 

[Rootax]

So, I played a few hours of sniper elite 4 on my pc (5820k@4.2ghz, 16gb of ram, and a fury X), @1440p, everything ultra and post processing FSAA at high. My video ram is full (It's always between 3900mb and 4016mb), but no stuttering at all, which is kind of nice, very smooth. I use DX12. So I don't know if AMD made a good driver with the 17.2.1, or if the game engine is pretty well done in swapping in and out stuff from vram without causing "pauses". But it's nice to see.

 

[Alessio1989]

I don't have the last Biohazard... does anyone have some information about it's streaming techniques or if at least if the game uses tiled resources/sparse textures/virtual textures/whatever_you_like_call_it?

As for Fallout 4... I would not take it in consideration... Except for a bug-race..