Support NeoGAF

[nemiroff]

I erased that part because i though i was coming out too confrontational

No worries, i like you, i can take it

 

[Abysse]

How did the RTX 2070 perform?

 

[Senua]

How did the RTX 2070 perform?

 

[Magic Carpet]

The overlay was reporting preset M while I played last night. Does that mean 4.5 is available to everyone? Or did I somehow accidentally enable developer mode or preview mode or alternative release mode?

 

[Thebonehead]

DLSS 4.5 is awesome for Star Citizen in VR. I don't have any screenshots to share here because it's VR but it is noticeably sharper in the headset with less artifacts and better performance with DLAA.

Is that a mod? Might have to reinstall it.

 

[DirtInUrEye]

This is kind of a shit show at the moment. Setting Nvidia App to latest version as a "set it and forget it" setting no longer seems viable.

Think I'm going to stick with Preset K (DLSS 4) for now, as I either run Quality or DLAA. L and M (DLSS 4.5) are Ultra Performance and Performance settings.

For anyone wondering what each preset does:

- Preset F (intended for Ultra Perf/DLAA modes): Will be deprecated in the next SDK cycle and should
not be used.
- Preset G (Unused): Do not use – reverts to default behavior
- Preset H (reserved): Do not use – reverts to default behavior
- Preset I (reserved): Do not use – reverts to default behavior
- Preset J: Similar to preset K. Preset J might exhibit slightly less ghosting at the cost of extra flickering.
Preset K is generally recommended over preset J
- Preset K: Default preset for DLAA/Balanced/Quality modes. Less expensive performance wise
compared to Preset L.
- Preset L: Default preset for UltraPerformance mode. Delivers a sharper, more stable image with less
ghosting than Preset J, K but are more expensive performance wise. Preset L is peak performant on
RTX 40 series GPUs and above.
- Preset M: Default preset for Performance mode. Delivers similar image quality improvements as
Preset L but closer in speed to Presets J, K. Preset M is peak performant on RTX 40 series GPUs and
above.

I'm considering downgrading to the previous GE so that whenever I select latest it can only go as high as 4.0.

I feel like there's a bit of good ol planned obsolescence being filtered through to us with these DLSS updates.

 

[Gitaroo]

so if I replace the dll with the new dlss 4.5 dlss the game should be using the transformer model? There are a few games that can't be swapped with nvidia app ie Alan Wake 2

 

[FalconPunch]

M Performance looks sharper than K Quality here. And has more fps. Is this witchcraft?

Based on Nvidia's own numbers, M having more fps than K should not be possible. So I'm going to guess that this dude made a mistake in his settings.

 

[kevboard]

Based on Nvidia's own numbers, M having more fps than K should not be possible. So I'm going to guess that this dude made a mistake in his settings.

not at the same preset, but this is comparing performance mode vs quality mode. so it's running at less than half the resolution in the M preset than in the K preset.

 

[dolabla]

On the Last of Us, preset M performance mode seems to look on par with native 4K and sometimes appears more sharper to my eyes. Performance mode may be the best way moving forward for me. I mean the picture quality is excellent and you get that boost in fps. This is so cool and will no doubt extend the life of these cards. Hadn't tried it yet on LoU, but I turned on smooth motion in the nvidia app since it doesn't have DLSS frame gen (only FSR) and it seems to work well.

 

[yamaci17]

1440p dlss performance ghosting and shimmering comparison



in this order,

preset D (ray reconstruction)
preset E
preset K
preset M

preset K introduces ghosting in the katana and preset M fixes it
both preset E and preset K has shimmering in road barriers and preset M fixes it
overall preset M looks the cleanest

 

[GymWolf]

I'm considering downgrading to the previous GE so that whenever I select latest it can only go as high as 4.0.

I feel like there's a bit of good ol planned obsolescence being filtered through to us with these DLSS updates.

Just use k profile if you want 4.0.

 

[DirtInUrEye]

Just use k profile if you want 4.0.

It's more that I can't be bothered having to keep checking to see what the latest letter recommendation is to make sure 4.0 is selected.

 

[Bojji]

It's more that I can't be bothered having to keep checking to see what the latest letter recommendation is to make sure 4.0 is selected.

Just force K as global profile and that's it.

 

[Tarnpanzer]

Just force K as global profile and that's it.

I heard this sometimes does not work for specific games.

 

[Bojji]

I heard this sometimes does not work for specific games.

It won't work for games not supported by Nvidia app (override). For that games use nvidia inspector or DLSS Swapper.

 

[nemiroff]

Is that a mod? Might have to reinstall it.

Yeah, it's awesome. But you might want to hold on a patch first. As usual CIG left for christmas vacation with the game in shambles after pushing a new big patch. The patch will probably arrive shortly.

And no, it's not an additional mod, it's baked into the releases.

It's demanding, so you'll need that DLSS...

 

[skneogaf]

So K is still best for dlaa or is M now the overall best for quality?

I play at 4k@120hz everything on absolutely maximum graphics settings.

Dlaa if my 5090 can give me enough fps.

 

[Imtjnotu]

So I switch my laptop from 2560x1600p to 1440 and set the present to M on dlss swapper.

Tested on BF6
Game still runs on ultra at 95-120 fps.

Will test my 4080S when I get home by switching the resolution to 1440p from 4K

 

[yamaci17]

1440p dlss performance ghosting and shimmering comparison



in this order,

preset D (ray reconstruction)
preset E
preset K
preset M

preset K introduces ghosting in the katana and preset M fixes it
both preset E and preset K has shimmering in road barriers and preset M fixes it
overall preset M looks the cleanest

similar comparison

1440p dlss quality preset E
1440p dlss quality preset K
1440p dlss balanced preset M
1440p dlss performance preset M

shimmering still happens with preset E and preset K at 1440p dlss quality

i will try ac shadows. that game had a lot of big problems with dlss 4.

 

[BennyBlanco]

If I install the new driver and keep the global settings to "use later version of DLSS" this will force 4.5?

 

[Buggy Loop]

If I install the new driver and keep the global settings to "use later version of DLSS" this will force 4.5?

Yes

 

[Buggy Loop]

Kind of a wasted video imo. Maybe Tim can make a better one later.
Quality vs quality is not what peoples are looking for. Talking about image quality, what it improves, performance 4.5 vs quality 4.0? Quality 4.5 vs DLLA 4.0?

 

[winjer]

Kind of a wasted video imo. Maybe Tim can make a better one later.
Quality vs quality is not what peoples are looking for. Talking about image quality, what it improves, performance 4.5 vs quality 4.0? Quality 4.5 vs DLLA 4.0?

Yes, Tim is already doing the quality comparison tests.
But the performance tests are also very important.

 

[Buggy Loop]

Yes, Tim is already doing the quality comparison tests.
But the performance tests are also very important.

I preferred the format they did in the past for performance tests

Yes it doesn't compare GPU generations but to me this is more interesting. I can see DLSS 3 vs DLSS 4 performances. So say, ok, DLSS 4 balanced ~ DLSS 3 quality in performance, now is the image quality of 4's balanced better than 3's quality? etc.

Anyway, Tim's video will probably do that format.

 

[T4keD0wN]

similar comparison

1440p dlss quality preset E
1440p dlss quality preset K
1440p dlss balanced preset M
1440p dlss performance preset M

shimmering still happens with preset E and preset K at 1440p dlss quality

i will try ac shadows. that game had a lot of big problems with dlss 4.

Have you tried preset L in quality/balanced/perf res scales to see how the image quality fares against M and K?

 

[yamaci17]

Have you tried preset L in quality/balanced/perf res scales to see how the image quality fares against M and K?

I didn't notice much of a difference compared to preset M with performance mode

 

[FIREKNIGHT2029]

Is that a mod? Might have to reinstall it.

No it's not a mod. Star Citizen Alpha 4.5 has official native VR support now and it's really good.

 

[3liteDragon]

winjer I see NVIDIA emphasizing how 4.5 uses the FP8 accelerators in the 40/50 series fully, does that mean 4.0 wasn't fully FP8? I was under the impression that it was.

 

[winjer]

winjer I see NVIDIA emphasizing how 4.5 uses the FP8 accelerators in the 40/50 series fully, does that mean 4.0 wasn't fully FP8? I was under the impression that it was.

Ampere and Turing used Int8 and Int4. And FP16.
This is why FSR4.5 is so slow, as it has to use FP16 for one calculation, while Ada can do 2 calculations of FP8.

 

[forsakkenSoul]

'This Really Puts a Nail in AMD FSR Redstone's Coffin': Clair Obscur: Expedition 33 NVIDIA DLSS 4.5 Performance Mode Stuns With Image Quality Close To Native 4K

Even in its early state, NVIDIA DLSS 4.5 stuns by delivering image quality better than native resolution, as in Clair Obscur: Expedition 33

wccftech.com

 

[Pop]

'This Really Puts a Nail in AMD FSR Redstone's Coffin': Clair Obscur: Expedition 33 NVIDIA DLSS 4.5 Performance Mode Stuns With Image Quality Close To Native 4K

Even in its early state, NVIDIA DLSS 4.5 stuns by delivering image quality better than native resolution, as in Clair Obscur: Expedition 33

wccftech.com

On my 5070 Preset M 4.5 performance at 4K is not only sharper and cleaner looking than Preset K 4.0 quality, also 4.5 fixes most of the ghosting problem of 4.0. And a 10 fps increase

I'll take it

 

[3liteDragon]

Ampere and Turing used Int8 and Int4. And FP16.
This is why FSR4.5 is so slow, as it has to use FP16 for one calculation, while Ada can do 2 calculations of FP8.

Was Ada using FP8 with 4.0?

 

[winjer]

Was Ada using FP8 with 4.0?

I think it was using Int8. Which is quite impressive, as it still looked better than FSR4 FP8. And way better than FSR4 Int8.

 

[3liteDragon]

I think it was using Int8. Which is quite impressive, as it still looked better than FSR4 FP8. And way better than FSR4 Int8.

Oh shit, I swear I thought Alex DF implied FP8 was being used. I thought the perf penalty for the older cards on 4.0 was bc of FP8

 

[Zathalus]

Pretty sure DLSS 4.0 always used FP8. Just that the new model is so heavy the lack of native FP8 on 2000 and 3000 series cards are more apparent. The old CNN model was INT8 I believe.

Nvidia mentions it here:

DLSS 4: Transforming Real-Time Graphics with AI

 

[PaintTinJr]

Ampere and Turing used Int8 and Int4. And FP16.
This is why FSR4.5 is so slow, as it has to use FP16 for one calculation, while Ada can do 2 calculations of FP8.

I always thought there was not such a doubling benefit on Nvidia CUDA cores and that on Ada it was the same FLOP count for FP16 as FP32 - and assumed that also was the same for FP8, and I'm sure it was mentioned as a downside for UE5 nanite on Nvidia versus AMD.

But given this is ML AI, I assume those are done in the tensor cores? and the flops double inside tensor cores with lower precision, presumably?

 

[Holammer]

Decided to play some Spongebob: Titans of the Tide. Used to play it at full epic settings @ 1440/240 with DLSSFG. Now it dropped to 200-210 (ugh! unplayable!).
Probably because there's no in-game option to select anything lower than DLSS Quality, switching back to Preset K fixed it.

 

[amigastar]

I've heard the final version comes out on the 13th, is that correct, i think i will wait for that one.

 

[Senua]

Kind of a wasted video imo. Maybe Tim can make a better one later.
Quality vs quality is not what peoples are looking for. Talking about image quality, what it improves, performance 4.5 vs quality 4.0? Quality 4.5 vs DLLA 4.0?

Rushed out to get initial hype views, fosho

 

[yamaci17]

damn, it looks the new model improves ac shadows a lot as well




preset M does not reduce rain particle density
preset M does not exhibit the weird ghosting that happens at the corner of the screen when you move the camera in specific direction
preset M does not exhibit the ghosting in flying particles and leaves
preset M does not seem to exhibit this weird artifacting that ruined some of my screenshots

performance figures and static comparisons

quality mode

https://imgsli.com/NDQxMDAy

balanced mode

https://imgsli.com/NDQxMDAz

performance mode

https://imgsli.com/NDQxMDA0

matched performance comparison on an rtx 3070 (preset K quality vs. preset M balanced)

https://imgsli.com/NDQxMDA1/1/2

moving comparisons (focus on tree branches)

https://imgsli.com/NDQxMDA5

https://imgsli.com/NDQxMDA5/1/3

https://imgsli.com/NDQxMDA5/2/6

https://imgsli.com/NDQxMDA5/1/3

none of the presets can fix the tree branches when using upscaling sadly, only native fixes it (for 1440p of course. 4k dlss performance looks much better). preset M seems to reduce the subtle ghosting that happens in tree branches though (this doesn't happen with TAA and preset E)

overall preset M provide great improvements and I'd prefer it over preset K on my system (by dropping from quality to balanced, I get roughly the same performance. but performance mode also looks decent).

 

[analog_future]

damn, it looks the new model improves ac shadows a lot as well




preset M does not reduce rain particle density
preset M does not exhibit the weird ghosting that happens at the corner of the screen when you move the camera in specific direction
preset M does not exhibit the ghosting in flying particles and leaves
preset M does not seem to exhibit this weird artifacting that ruined some of my screenshots

performance figures and static comparisons

quality mode

https://imgsli.com/NDQxMDAy

balanced mode

https://imgsli.com/NDQxMDAz

performance mode

https://imgsli.com/NDQxMDA0

matched performance comparison on an rtx 3070 (preset K quality vs. preset M balanced)

https://imgsli.com/NDQxMDA1/1/2

moving comparisons (focus on tree branches)

https://imgsli.com/NDQxMDA5

https://imgsli.com/NDQxMDA5/1/3

https://imgsli.com/NDQxMDA5/2/6

https://imgsli.com/NDQxMDA5/1/3

none of the presets can fix the tree branches when using upscaling sadly, only native fixes it. preset M seems to reduce the subtle ghosting that happens in tree branches though (this doesn't happen with TAA and preset E)

overall preset M provide great improvements and I'd prefer it over preset K on my system (by dropping from quality to balanced, I get roughly the same performance. but performance mode also looks decent).

We're at the point with DLSS 4.5 where Quality upscaling is legitimately better than native 4k with virtually no drawbacks. Awesome stuff.

 

[Gitaroo]

On the Last of Us, preset M performance mode seems to look on par with native 4K and sometimes appears more sharper to my eyes. Performance mode may be the best way moving forward for me. I mean the picture quality is excellent and you get that boost in fps. This is so cool and will no doubt extend the life of these cards. Hadn't tried it yet on LoU, but I turned on smooth motion in the nvidia app since it doesn't have DLSS frame gen (only FSR) and it seems to work well.

that's how I am playing the game, Smooth motion for older games without native frame gen support works almost like native DLSS 3 FG which it kind of it since it repurpose the optical flow accelerator for that task. Input lag is barely noticeable in 100+ fps. I use it for all fighting games even and 120fps fighting games are smoooooth.

 

[daninthemix]

Anyone done any 4.5 vs PSSR comparisons yet?

Oh wait, no point.

 

[Senua]

Anyone done any 4.5 vs PSSR comparisons yet?

 

[Bojji]

Anyone done any 4.5 vs PSSR comparisons yet?

Oh wait, no point.

PSSR was losing to DLSS2/3. They did fucking nothing to improve it in 14 months (we had two transformer models released in that time and FSR4)...

 

[xVodevil]

So I had to check 4K Ultra performance + L to see for myself. This being actually as good as it is in some of the games is kinda wild, not always, but damn.

 

[GymWolf]

On my 5070 Preset M 4.5 performance at 4K is not only sharper and cleaner looking than Preset K 4.0 quality, also 4.5 fixes most of the ghosting problem of 4.0. And a 10 fps increase

I'll take it

4.5 performance looks better than K quality? For real?

 

[GymWolf]

If I install the new driver and keep the global settings to "use later version of DLSS" this will force 4.5?

You have to update the app aswell.

 

[sertopico]

Just tested CP 2077:

4090
3440x1440
DLSS Q latest model wit res scaling @85%

It definitely further improves over the older model in many areas where the version 4 was still lacking. Ghosting got better, in particular on those little screens all around the city with some scrolling text visible (like on the sidewalks). In the previous versions staring at those without moving the camera for a while started making the text all blurry, now the issue is completely gone on my end. Aliasing got also better and overall distant objects are much sharper. It almost makes sharpening superfluous. I am impressed.

I also tested it on Jedi Survivor which I bough recently due to the Steam sales, but unfortunately any model different from the original DLSS 3 causes some very annoying bugs on the vegetation in the shadows which "boils" and shimmers. Pity!