ada lovelace NVIDIA Project Beyond GTC Keynote with CEO Jensen Huang: RTX 4090 + RTX 4080 Revealed

[porina]

Looks like I totally misunderstood DLSS 3 operation. Watching Digital Foundry's video on it now. So it is interpolation, not extrapolation. That does get around one question I had if it were extrapolation, which is how do you handle data you don't have? Interpolation doesn't have to worry so much about that. Implied latency increase then is about half the frame time? So at 60fps native rendering, that's about +8ms cost, but this gets worse if you have lower native fps which is where more of the bonus would be. The saving point for it seems to be if you combine DLSS 2 that decreases latency more than the increase from DLSS 3, so net reduction from native rendering.

[HenrySalayne]

51 minutes ago, porina said:

Watching Digital Foundry's video on it now. So it is interpolation, not extrapolation. That does get around one question I had if it were extrapolation, which is how do you handle data you don't have? Interpolation doesn't have to worry so much about that. Implied latency increase then is about half the frame time? So at 60fps native rendering, that's about +8ms cost, but this gets worse if you have lower native fps which is where more of the bonus would be.

The "inter" in interpolation means "in-between". So you would need two adjacent frames and the delay would be 1 frame if the computation would be instant (and to much surprise to nobody - it's not instant).

Frame N and Frame N+1 need to be available to interpolate frame N+0.5 . So this is pretty much what your good old mid 2010s tellies do. Just slightly more sophisticated.

[porina]

11 minutes ago, HenrySalayne said:

The "inter" in interpolation means "in-between". So you would need two adjacent frames and the delay would be 1 frame if the computation would be instant (and to much surprise to nobody - it's not instant).

Frame N and Frame N+1 need to be available to interpolate frame N+0.5 . So this is pretty much what your good old mid 2010s tellies do. Just slightly more sophisticated.

Ok, I neglected the computation time in my previous post. The DF video was talking of the magnitude of several ms, so depending on the real frame generation rate it could be a notable proportion of that.

 

It is arguably a lot more sophisticated. DF's video also compared it against two other frame rate increasers from Adobe and Topaz. DLSS 3 was clearly working better, likely from it having the motion data from the game engine to work with and not just working it out from the output images.

 

Ultimate low latency gamers would probably still play at low resolutions without any extra processing. Especially at higher resolutions, if you compare to native, using combo of DLSS 2 + 3 is still an overall latency reduction.

[starsmine]

So... its interpolation rather then extroplation and it forces reflex on to mitigate


SO it uses dlss 2.0 which is how it beats native
buffers a frame so in imaginary land where reflex was off latency would be in the 60 to 70s who knows
And reflex puts it down between reflex on and off of dlss 2.0

I think I would rather just use dlss 2 rendering with reflex and some form of freesync in 90% of cases.
Cause here dlss 3.0 is worse then 2.0 with reflex off, but still way better then native

[Quackers101]

DF released video on DLSS 3, some as seen from above^

If not mistaken intel could do this too, only needs to put it together and might have to adjust to add this for battlemage.

as for the AI options and some of the data flow and how they use some similar data for XeSS, but needs the generation etc. If the pipeline works for something that gives any decent result or speed.

Edited September 29, 2022 by Quackers101

[HenrySalayne]

7 hours ago, starsmine said:

So... its interpolation rather then extroplation and it forces reflex on to mitigate


SO it uses dlss 2.0 which is how it beats native
buffers a frame so in imaginary land where reflex was off latency would be in the 60 to 70s who knows
And reflex puts it down between reflex on and off of dlss 2.0

I think I would rather just use dlss 2 rendering with reflex and some form of freesync in 90% of cases.
Cause here dlss 3.0 is worse then 2.0 with reflex off, but still way better then native

 

Could somebody explain these numbers to me? How is rendering frame N and displaying it slower than waiting for frame N+1 to render frame N+0.5 and displaying it? Especially the minuscule difference in Portal Remastered compared to DLSS 2 doesn't make any sense.

Everbody explains DLSS 3 as a fancy intermediate image rendering. But this would mean a minimum of 1 frame of latency which we should not see with DLSS 2 or native rendering.

[starsmine]

47 minutes ago, HenrySalayne said:

 

Could somebody explain these numbers to me? How is rendering frame N and displaying it slower than waiting for frame N+1 to render frame N+0.5 and displaying it? Especially the minuscule difference in Portal Remastered compared to DLSS 2 doesn't make any sense.

Everbody explains DLSS 3 as a fancy intermediate image rendering. But this would mean a minimum of 1 frame of latency which we should not see with DLSS 2 or native rendering.

The latency pipeline is not just the frame rendered.
DLSS 3.0 probably does not even wait for the full frame to render before it starts interpolating, The tensor cores that are doing nothing can probably just yoink most of the pre processing data it needs before frame n+1 is finished. DLSS 3.0 is not interpolation in the classical sense. it does not appear to just be taking two frames and averaging them together, but rather its taking data from the rendering pipeline and making a guess on where it was in right before it. 


So traditional rendering with reflex off in the above graph. When the display is showing frame 1, the CPU is already done with frame 4 that data can likely be pre processed by the tensor cores before making the final calculations when the GPU finished frame 4 and can insert a 3.5 frame before it. 

But im just speculating. 

 

[porina]

5 hours ago, HenrySalayne said:

Could somebody explain these numbers to me? How is rendering frame N and displaying it slower than waiting for frame N+1 to render frame N+0.5 and displaying it?

Keep in mind frame times themselves, rather than just counting frames. One way of looking at it is that DLSS 2 is used to reduce the frame time compared to native, and DLSS 3 uses some of that saving to generate the extra frames. If you skip the DLSS 2 part, then DLSS 3 would increase latency relative to native.

[Quackers101]

6 hours ago, HenrySalayne said:

Could somebody explain these numbers to me? How is rendering frame N and displaying it slower than waiting for frame N+1 to render frame N+0.5 and displaying it? Especially the minuscule difference in Portal Remastered compared to DLSS 2 doesn't make any sense.

Everbody explains DLSS 3 as a fancy intermediate image rendering. But this would mean a minimum of 1 frame of latency which we should not see with DLSS 2 or native rendering.

You take the data used in DLSS 2 and do more with it. so you already half of the info from previous frame then the rest with the newer frame.

A bit how intel does things in parallel for their XMX, motion vectors, resolution etc. Got no idea on the AI part, which I guess one of the creators have maybe talked a bit about on social media. would assume it goes through the same check again like with the other images to not give more artifacts but artifacts from DLSS 2 could be worse or better depending on what it does? (as seen with some being worse in DLSS 3 or better than DLSS 2 in DF's video).

Also how they can know what data to use vs what to throw away, so you don't need to do the full image every time.

Edited September 29, 2022 by Quackers101

[hamilton21017]

https://videocardz.com/newz/nvidia-geforce-rtx-4090-performance-showcased-in-digital-foundrys-first-look-at-dlss-3-technology

DLSS3 ON→Reflex ON→+20ms
For some games, using DLSS3 seems to increase the input delay by +20ms or more.

 

I feel this would be fatal in a competitive game.

[MageTank]

32 minutes ago, hamilton21017 said:

https://videocardz.com/newz/nvidia-geforce-rtx-4090-performance-showcased-in-digital-foundrys-first-look-at-dlss-3-technology

DLSS3 ON→Reflex ON→+20ms
For some games, using DLSS3 seems to increase the input delay by +20ms or more.

 

I feel this would be fatal in a competitive game.

Most competitive titles are not graphically intense though. You can get Rainbow Six Siege to run on a potato, same with Overwatch and CS:GO. This is ignoring games like League that can be ran on carrier pigeons carrying papers with one's and zero's written on them.

 

For AAA RPG games that focus on "realism" and maximum immersion, this is probably going to be considered a net win. Games like Cyberpunk, whatever the new Skyrim game is called, and other casual story-driven games, those will the ones that benefit the most from DLSS3.

[Dracarris]

43 minutes ago, hamilton21017 said:

I feel this would be fatal in a competitive game.

Can we please stop this ridiculous latency discussion? If you care about a few ms of additional latency and successfully pretend they will actually make a difference, well then turn DLSS off. Does it matter whether you will play CSGO at 650 or 500fps?

 

This feature is clearly for eyecandy games where the full raw power of the GPU is not enough to reach a stable 60-100fps experience. Which can happen easily at raytraced, detailed games at 4K. Doing competitive multiplayer? Well turn it off and go for 1080p high settings or whatever. Pew Pew.

[ewitte]

On 9/20/2022 at 4:49 PM, Shimmy Gummi said:

They do NOT list core count. A typical user would look at these and come to the conclusion that other than the memory difference, they are the same. Lucky for them, the difference is marginal, albeit present.

 

I wonder how likely the 4080 12gb and 16gb will have similar marketing material?

Its the same core and for such a small difference it is fine in the instance of the 3080 12GB.  The issue is when you do something that drastically changes the product like the "4080 12GB" or changing to DDR3 on the GT 1030.  Or releasing a SSD to reviewers and then gimping it after the reviews come out...

[ewitte]

15 minutes ago, Dracarris said:

Can we please stop this ridiculous latency discussion? If you care about a few ms of additional latency and successfully pretend they will actually make a difference, well then turn DLSS off. Does it matter whether you will play CSGO at 650 or 500fps?

 

This feature is clearly for eyecandy games where the full raw power of the GPU is not enough to reach a stable 60-100fps experience. Which can happen easily at raytraced, detailed games at 4K. Doing competitive multiplayer? Well turn it off and go for 1080p high settings or whatever. Pew Pew.

The latency numbers with reflex on seem fine even better than native in some instances.

 

 

[Quackers101]

2 hours ago, Dracarris said:

Can we please stop this ridiculous latency discussion? If you care about a few ms of additional latency and successfully pretend they will actually make a difference, well then turn DLSS off. Does it matter whether you will play CSGO at 650 or 500fps?

 

This feature is clearly for eyecandy games where the full raw power of the GPU is not enough to reach a stable 60-100fps experience. Which can happen easily at raytraced, detailed games at 4K. Doing competitive multiplayer? Well turn it off and go for 1080p high settings or whatever. Pew Pew.

also in PvP or some competitive styles, its just not about only the "latency", but also what information the image sends to the player, which can mean some image quality/fidelity. Something DLSS 3 improves in some movement, but worse in other movements like when ghosting or lack of data for reconstruction then brought into image generation. So the future of DLSS 3 can be for both, but some might want to turn it off.

1 hour ago, ewitte said:

Its the same core and for such a small difference it is fine in the instance of the 3080 12GB.  The issue is when you do something that drastically changes the product like the "4080 12GB" or changing to DDR3 on the GT 1030.  Or releasing a SSD to reviewers and then gimping it after the reviews come out...

oof the SSD switch, also that lower memory width bit per card, could maybe hinder DLSS 3 performance by all the data it has to process. As the cards used like the 4090 in DF's video and how using the 3090 generation. so wonder how the low end 4080 does with DLSS 3, or if the memory 384 bit vs 192 bit matters or just the cores, speed etc?

[ewitte]

23 hours ago, Shimmy Gummi said:

This is pretty much my limiting factor. Our home office is 15a circuit, has two computer systems on it, along with some other LED home lighting (not much power, but still). My understanding is you don't want to push more than 80% sustained on the circuit's rated load, which for 15a would be 1800w? So ~1400w, let's say 1200w when you take into account all the other lighting on the circuit.

 

That means both systems should not pull more than 600w continuous without possibly damaging the wiring. 

 

Now, most gaming sessions my PC's power strip completely on a killowatt meter is around 600w........that's a 10900KF OC + 3080 and 3 displays + a fan and speakers. Wife' system is less powerful, 9900k stock, 3070, 3 displays and speakers. 

 

Now, double that, and I'm already at 1200w continuous. If games start using more CPU load, then that's 1400w already.

 

I can't upgrade to a more powerful GPU on either PC unless I upgrade to a 20a circuit. I even moved our wifi Brother laser printer to another circuit, because it can pull over 400w when it fires up.

I'll likely put dedicated lines in the next house for the office, possibly 240v for the computer(s).  The secondary machine is pretty efficient 200W max for a 5600g and 3060.

[hamilton21017]

Call of Duty® | Modern Warfare II

 

In my opinion, the DLSS3 delay would be fatal in a fast kill time FPS where you are competing against each other for tenths of a second during a firefight.

 

[Dracarris]

6 minutes ago, hamilton21017 said:

In my opinion, the DLSS3 delay would be fatal in a fast kill time FPS where you are competing against each other for tenths of a second during a firefight.

How large is the network latency of such a setup?

[starsmine]

8 minutes ago, hamilton21017 said:

Call of Duty® | Modern Warfare II

 

In my opinion, the DLSS3 delay would be fatal in a fast kill time FPS where you are competing against each other for tenths of a second during a firefight.

 

you are talking in the order of .1 seconds when the rest of us are talking in the order of .001 seconds mate.
What is that link even showing us?
And what are you comparing dlss3 to? Because even WORST CASE that has been tested as of yet, the spidyman game, dlss 3.0 was 38ms vs native reflex off was 39ms, native with it on was 36ms.

If you are fighting for tenths of a second, this which is 100ms, and dlss 3.0 adds 2ms over native. You focused on the wrong thing mate.

But also, why you playing the game at 4k anyways is the real question to be asking if you care that much about it.
Mis use a tool of course things will go south.

[HenrySalayne]

6 hours ago, Dracarris said:

Can we please stop this ridiculous latency discussion? If you care about a few ms of additional latency and successfully pretend they will actually make a difference, well then turn DLSS off. Does it matter whether you will play CSGO at 650 or 500fps?

 

This feature is clearly for eyecandy games where the full raw power of the GPU is not enough to reach a stable 60-100fps experience.

I don't think you understand. It's not about 5 ms, it's about a full frame (or even more) especially when fps are pretty low to begin with. If you reach 40 Hz native these additional 25 ms of delay to wait for frame N+1 are noticeable. Even the buttery smooth picture won't fool you about the input latency.

 

2 hours ago, Dracarris said:

How large is the network latency of such a setup?

How bad is 10 ms of audio delay? Not too bad. How bad is 10 ms of audio delay on a single device? Terrible! Latency adds up across the chain and every single part needs to be as low as possible.

[pas008]

2 hours ago, hamilton21017 said:

Call of Duty® | Modern Warfare II

 

In my opinion, the DLSS3 delay would be fatal in a fast kill time FPS where you are competing against each other for tenths of a second during a firefight.

 

why would you be using dlss3 on pvp game in first place, most lower settings in first place to make sure you maintain dips

[Spicedaddy]

So DLSS 2 is upsampling and DLSS 3 is frame interpolation. I had this 10 years ago, playing at lower resolutions on a "120Hz MotionFlow" TV. 

 

What's next? Back to black and white! Colors are barely noticeable anyways, and you get 4X the FPS.

 

Seriously nVidia, make the hardware better instead of cutting corners.

[BiG StroOnZ]

Newegg has some AIB/AIC listings for the RTX 4090:

 

 

[Shzzit]

HOW the hell is a strix 400 dollars OVER msrp?  You can get tha AIO one for way less lol.

 

What makes it worth 2000?  What a 20mhz OC bios?  /facepalm.

 

Thats the one i wanted too.

 

 

Guess my best bet for EK waterblock is just getting an FE right?  FE have x2 bios?

[IPD]

Oh look, it's a Titan 2.  Just like I said the pricing would be.