NVidia Ada Speculation, Rumours and Discussion

[troyan]

Rasterizing is an outdated concept with these APIs. Looking at the architecture the 4090 has nearly twice of everything over the 3090 except bandwidth:
2x rasterizing, 2x geometrie performance, 2x FP32 performance

Yet in Valhalla the GPU is only 1.65x faster which would be only 2.4x faster than the 3070.

 

[techuse]

If I was buying Ada the only card I would consider is the 4090. The 4080 12GB Is a joke both in name and performance. I’m not drawing any final conclusions until we see 3rd party benchmarks in realistic scenarios without the DLSS performance mode nonsense.

Even DFs teaser benchmarks were for 4K DLSS performance. They never did that for the 3090 so why do it for the 4090? Silly.

Probably all Nvidia would allow at this point.

 

[Kaotik]

The question now is: with the crazy tripling of the transistor budget, where did all of this budget go? Towards RT cores and Tensor cores? The 4090 probably needs more power and more clocks to push towards 2X the rasterization that was expected from it. Either that or the architecture is not as scalable as it used to be.

96MB of L2 eats a bunch. That's 16x GA102

 

[Below2D]

These are HPC cards NVIDIA is selling as gaming cards. This massive compute performance uplift doesn’t scale well with gaming at all.

 

[DegustatoR]

These are HPC cards NVIDIA is selling as gaming cards. This massive compute performance uplift doesn’t scale well with gaming at all.

Could've sworn I've heard this two years ago... And then there were no better gaming cards on the market still.

 

[DavidGraham]

Even DFs teaser benchmarks were for 4K DLSS performance. They never did that for the 3090 so why do it for the 4090? Silly.

It's something of an integrity thing on the part of NVIDIA actually.

You see, DLSS 3 is going to work on the basis of DLSS2 in Performance mode, the first frame is made by DLSS2 Performance, the following frame will be done through DLSS3 (Frame Generator). So comparing against DLSS2 Performance on the 3090Ti is a must here. This way NVIDIA isolated the only new variable in the new RTX 4000 series, the frame generation part.

If NVIDIA compared against native or DLSS2 Quality on the 3090Ti, the gains would be astronomically high, but also misleadingly high as well.

 

[techuse]

It's something of an integrity thing on the part of NVIDIA actually.

You see, DLSS 3 is going to work on the basis of DLSS2 in Performance mode, the first frame is made by DLSS2 Performance, the following frame will be done through DLSS3 (Frame Generator). So comparing against DLSS2 Performance on the 3090Ti is a must here. This way NVIDIA isolated the only new variable in the new RTX 4000 series, the frame generation part.

If NVIDIA compared against native or DLSS2 Quality on the 3090Ti, the gains would be astronomically high, but also misleadingly high as well.

I think he was wanting to see benchmarks at either native or a higher DLSS level without frame generation on Ada.

 

[Scott_Arm]

...

Creativity Redefined: New GeForce RTX 40 Series GPUs and NVIDIA Studio Updates Accelerate AI Revolution

Next-generation ray tracing and Tensor Cores plus dual AV1 NVIDIA encoders power new tools for 3D, video production, live streaming, game modding and more.

blogs.nvidia.com

...

Wtf is this graphic. I can't tell which card is which. I can assume 4090 is top, 3080 ti is bottom, but I can't tell the difference between the middle two most likely because of my colour blindess. Honestly no idea how websites etc are still absolutely brain damaged about designed graphics for the 5% of the world with some form of colour vision anomaly.

 

[Jawed]

HBM3 would have sorted the scaling...

 

[DegustatoR]

HBM3 would have sorted the scaling...

And doubled the prices.

 

[Deleted member 2197]

EK Water Blocks water block makes your GeForce RTX 4090 a single slot design

The EK Water Blocks' EK-Quantum Vector2 FE RTX 4090 D-RGB series of water blocks is compatible with GeForce RTX 4090 Founders Edition....

www.guru3d.com

 

[Jawed]

And doubled the prices.

Which makes no difference to the people buying 4090

 

[techuse]

Which makes no difference to the people buying 4090

Agreed. Anyone who would buy one of these GPUs doesn't care at all about any semblance of value and would probably prefer the price be higher if it means better performance.

 

[Malo]

Which makes no difference to the people buying 4090

Or 4080 something Gb.

 

[neckthrough]

Sentiment around the 4080s appears to be biasing objective critique of the 4090. The two segments have very different narratives and should be critiqued independently. Keep in mind that the 4090 has a 24GB frame buffer and is bringing a sizeable generational improvement vs. the 3090 while costing less (inflation-adjusted).

 

[DegustatoR]

Sentiment around the 4080s appears to be biasing objective critique of the 4090. The two segments have very different narratives and should be critiqued independently. Keep in mind that the 4090 has a 24GB frame buffer and is bringing a sizeable generational improvement vs. the 3090 while costing less (inflation-adjusted).

4080 16GB has a 16GB of VRAM and likely brings similarly sizeable generational improvements over any Ampere part.
The only card which looks "weak" is 4080 12GB - but mostly because all Ampere lineup above 3080 10GB was condensed into a +-20% of performance. Which makes 4080 12GB vs 3090Ti performance comparisons as relevant as 4080 12GB vs 3080 10GB really.

 

[DegustatoR]

NVIDIA Ada AD102 Block Diagram and New Architectural Features Detailed

At the heart of the GeForce RTX 4090 is the gigantic AD102 silicon, which we broadly detailed in an older article. Built on the 4 nm silicon fabrication process, this chip measures 608 mm² in die-area, and crams in 76.3 billion transistors. We now have our first look into the silicon-level block...

www.techpowerup.com

 

[techuse]

4080 16GB has a 16GB of VRAM and likely brings similarly sizeable generational improvements over any Ampere part.
The only card which looks "weak" is 4080 12GB - but mostly because all Ampere lineup above 3080 10GB was condensed into a +-20% of performance. Which makes 4080 12GB vs 3090Ti performance comparisons as relevant as 4080 12GB vs 3080 10GB really.

Which Ampere GPU do you think is the proper comparison point for each Ada product?

 

[swaaye]

96MB of L2 eats a bunch. That's 16x GA102

I just remembered GA100. It had 40MB L2 cache.

NVIDIA Tesla A100 with GA100 Ampere GPU based on 7nm, 54b transistors and 6912 CUDA Cores

It won't be much longer before all details will be released and announced, but meanwhile there is a photo of the upcoming Tesla A100 SMX module, fitted with an GA100 GPU (a for Ampere). it shows HB...

www.guru3d.com

 

[Dictator]

It looks like they've refurbished the 3Dfx T-Buffer Motion Blur effect.

T Buffer effect Just overlayed previous internal frames with Increasing opacity, this looks really different