NVidia Ada Speculation, Rumours and Discussion

[Scott_Arm]

https://twitter.com/x/status/1572257012263657475

vewy not fast

The way they're releasing numbers is scummy. It should show which games are using DLSS3, DLSS2 or no DLSS at all. Clearly the three on the far right are DLSS3 supporting, and I think MS FS is as well. Interesting case there because Flight Sim is typically cpu limited, so all three new cards basically hit the same performance with DLSS 3. It's doing a nice job of increasing the fps under cpu limited scenario, but that also means the cheapest card does as well as the most expensive one.

Also, the 4080 12GB is a 285W card, so it's doing very well compared to a 3090ti. I wouldn't say they're not fast.

 

[Maillog]

Does it buffer up multiple frames to do the interpolation ?

Yes, interpolation is execute by a separate unit in the GPU
But Optical Flow SDK is used to "create" motion vectors in real time for an already completed image (e.g., a movie). How it will work in games is not clear yet

 

[DegustatoR]

Did you even watch the event? They claim 61 FPS for DLSS2 and 101 FPS for DLSS3 in Cyberpunks new RT mode

Did you?

 

[Scott_Arm]

 

[Jawed]

Ideas for new uses by NVidia for tensor cores:

1. Frame doubler, using AI-driven interpolation. Many games use a G-buffer and often there are motion vectors associated with G-buffers. In general, I presume, G-buffers are fully complete less than half-way through the time it takes to render a frame. This would enable an inference engine to take two frames' G-buffers, motion vectors (if available) and render targets to interpolate a third frame (underlined in the list below) while the fourth frame is being fully rendered (turning 60fps rendering into 120fps display):
   • 8.333ms: frame 1 inferenced G-buffer, motion-vectors and render target
   • 16.67ms: frame 2 real G-buffer, motion-vectors and render target
   • 25.00ms: frame 3 inferenced G-buffer, motion-vectors and render target generated while frame 4 is being shaded
   • 33.33ms: frame 4 real G-buffer, motion-vectors and render target
2. Ray-generation optimiser. Generating the best rays to trace seems to be a black art. It's unclear to me how developers tune their algorithms to do this. I'm going to guess that with some AI inferencing that takes the denoiser input and some ray generation statistics, the next frame's ray generation can be steered for higher ray count per pixel where quality matters most. This would enable faster stabilisation for algorithms such as ambient occlusion or reflections where disocclusion and rapid panning cause highly visible artefacts due to low ray count per pixel.

I'm going to guess that these ideas would use a lot of memory. I get the sense that the next generation is going to be more generous with memory, so that shouldn't be a problem.

I'm expecting these kinds of ideas to make use of the DLL-driven plug-in architecture we've seen with DLSS, activated by developers as a new option in the graphics settings menus.

How much of this kind of stuff can be back-ported to Turing or Ampere? What would make these features uniquely available on Ada?

DLSS 3 to double fps instead of pixels ..

Yay I got number 1 "right".

It seems like it's too demanding for prior architectures.

 

[Scott_Arm]

@Jawed I'm guessing it's using the new fp8 format, which I don't think ampere supports.

 

[gamervivek]

nvidia still have only DP1.4a on these. And to reiterate, no mention of 8k gaming despite DLSS3 being a perfect tool to have 8k high refresh rate become a reality.

 

[Lurkmass]

Here's what one of the Portal Desolation developer had to say about Portal RTX ...

https://twitter.com/x/status/1572266254458621952

 

[DegustatoR]

nvidia still have only DP1.4a on these

You need more? DP1.4 w/DSC allows you to drive a 4K display at 240Hz.

And to reiterate, no mention of 8k gaming despite DLSS3 being a perfect tool to have 8k high refresh rate become a reality.

Thank god for that. We still don't really have good 4K monitors, and it will be some years until 8K will be of any interest.

 

[Phantom88]

They're showing an Asus card here. He says its over 36 cm long. A monster. It means it wont fit in a lot of cases. A meshify C, very popular case for example, has a clearance of 31 cm if i remember right with fans in front and 33 if you remove the fans. So, that case and other like it, buy a new case as well with your 2000 euro gpu

 

[gamervivek]

8k 120Hz shold be easily doable on these cards and 4k is already too low PPI on the smaller OLED TVs. Maybe nvidia are saving it for 50xx series, dunno wth they are thinking.

edit: the footnotes on the specs page do mention this though:

1 - Up to 4k 12-bit HDR at 240Hz with DP 1.4a + DSC. Up to 8k 12-bit HDR at 60Hz with DP 1.4a + DSC or HDMI 2.1 + DSC. With dual DP 1.4a + DSC, up to 8K HDR at 120Hz

 

[pjbliverpool]

Okay so no-where near 4x RT performance then. Cyberpunk maxed out (one of the heaviest RT tests available presently) is generally less than 2.5x faster when using DLSS3 on the 4090 vs DLSS2 on the 3090Ti. So probably a little under 70% faster with no DLSS (based on the DLSS2 vs DLSS3 test showing a roughly 50% uplift for 3). So fast, but below expectations and certainly slower than it was portrayed in the video. The 4080 12GB (4070) is probably going to struggle to keep up with the 3090Ti in many current games even with RT. Not bad if it were priced like a 4070 but not even close to worth the 80%(!) price increase for a performance level that we would generally expect from a new generation in this category.

 

[Jawed]

NVIDIA introduces DLSS 3 upscaling technology, 35 games on a list - VideoCardz.com

NVIDIA DLSS 3: AI-Powered Performance Multiplier Boosts Frame Rates By Up To 4X NVIDIA DLSS revolutionized graphics by using AI super resolution and Tensor Cores on GeForce RTX GPUs to boost frame rates while delivering crisp, high quality images that rival native resolution. Since the release...

videocardz.com

Building upon DLSS Super Resolution, DLSS 3 adds Optical Multi Frame Generation to generate entirely new frames, and integrates NVIDIA Reflex low latency technology for optimal responsiveness. DLSS 3 is powered by the new fourth-generation Tensor Cores and Optical Flow Accelerator of the NVIDIA Ada Lovelace architecture, which powers GeForce RTX 40 Series graphics cards.

The DLSS Frame Generation convolutional autoencoder takes 4 inputs – current and prior game frames, an optical flow field generated by Ada’s Optical Flow Accelerator, and game engine data such as motion vectors and depth.

More detail on the page.

 

[hurleybird]

Anyone notice that Nvidia claims 90TF for "Ada" but the 4090 only works out to 2 x 16384 x 2520 = 82.575TF? Scummy? Preview of 4090 Ti?

 

[troyan]

Anyone notice that Nvidia claims 90TF for "Ada" but the 4090 only works out to 2 x 16384 x 2520 = 82.575TF? Scummy? Preview of 4090 Ti?

Was rounded up. In the trailer it is 83 TFLOPs.

 

[DegustatoR]

NVIDIA Micromesh

A graphics primitive for micro triangles that can encode opacity or displacement.

developer.nvidia.com

So these aren't going to be included into DXR soon I take it?

Not bad if it were priced like a 4070 but not even close to worth the 80%(!) price increase for a performance level that we would generally expect from a new generation in this category.

3090-3090Ti were $1500-2500 though, 4080 12GB moves this into $900.
It is a sizeable perf/price improvement albeit the prices themselves are anything but mainstream.
4080 12GB is a weirdly positioned product wrt VRAM size though. I'd expect it to be able to do 4K+RT easily but that 12GB buffer may not be enough for that over the next couple of years.

Anyone notice that Nvidia claims 90TF for "Ada" but the 4090 only works out to 2 x 16384 x 2520 = 82.575TF? Scummy? Preview of 4090 Ti?

Real boost figures are always higher than specced on Nvidia GPUs. It will likely run at 90TF in practice meaning sustained clock at ~2.75GHz.

 

[Jay]

Really may end up being intresting months ahead.

We could be looking at:
Intel Arc - Low end / entry (obviously drivers withstanding)
AMD RDNA3 - Middle low - high (depending on RT perf, where it lands price etc)
Nvidia 4080+ - High end

 

[Scott_Arm]

@DegustatoR Yah, I'll worry about the pricing when I can actually see the performance relative to the 30 series with DLSS off etc. If you get 3090/3090 ti performance at $900 USD, that's a good deal. I'd like to see how the 4080 12GB compares to a 3080. I also figure that going forward DLSS support is going to be pretty pervasive as most games/engines are supporting motion vectors now. If that ends up being the case, the raw performance difference is going to be a lot smaller than the real practical performance difference with DLSS 3.

 

[Dampf]

As I understand it DLSS3 will use frame interpolation (similar to what you can see in SVP) and work with the Optic Flow SDK
And there is a block for acceleration starting from Turing and it works independently from the rest of the GPU

Wait what?! Is that the optical flow hardware that is described in the RTX 4000 marketing?

If that's the case and the hardware for DLSS3 is really in Turing and Ampere, then that's HUGE BS from Nvidia.

 

[davis.anthony]

This is kind of reminding me of the GTX 280/260 launch. Wow, the tech. Damn, the price.

There's an opening here for a 4870/4850 moment with a lower performance AMD card if they can get within sniffing distance of Ada performance at like half the cost similar to what happened with that generation.

It doesn't have to be faster, it'd just have to be somewhat close (say within 25-35% of Ada performance) and come in at a significantly cheaper price.

Regards,
SB

The HD4870 was real punch to the gut for Nvidia.

And the HD4890 after that too.

Another moment like that from AMD would be epic.