NVidia Ada Speculation, Rumours and Discussion

Status
Not open for further replies.
Read the footnotes. It says at 4k with DLSS performance mode. This is embarrassing from Nvidia. I’m not optimistic about the real raster performance improvements.

~70% faster in the Division 2, which has a good DX12 renderer unlike Valhalla...
 
Last edited:
Read the footnotes. It says at 4k with DLSS performance mode. This is embarrassing from Nvidia. I’m not optimistic about the real raster performance improvements.
1. RE8, ACV, TD2 do not support DLSS.
2. All of these are running at >100 fps on 3090 in 4K at max settings. I won't be surprised if they are CPU limited on 4090 in the same scenario - and that this slide shows what DLSS3 brings to such cases. MSFS is 100% CPU limited as can be seen from lack of scaling between all three cards.

Have we all already forgotten that we should wait for proper benchmarks?
 
~70% faster in the Devision 2, which has a good DX12 renderer unlike Valhalla...
Using DLSS performance. Frankly, DLSS performance improvements are “fake” improvements to me because, I don’t use DLSS due to the image errors. I’m interested to see what the non-DLSS performance improvements look like. 99% of games created don’t use DLSS and so DLSS performance is basically irrelevant to me at this time.
 
The pricing is interesting in that it's against typical convention in that instead of diminishing returns at the high end the RTX 4090 actually ends up at significantly better value than the 4080 12GB. This effectively results in the lack luster pricing structure further down the stack. RTX 4090 as such also ends up as the segment with the greatest value increase when compared to the RTX 3090 at launch.

DLSS 3 effectively at this point ends up as a bit a double edged sword for consumers. If Nvidia is confident that they can deliver DLSS 3 in terms of both function and messaging then this does make the RTX 4080 12GB enticing relative to the other options in the market by itself. The flip side is it does support the a price that otherwise would likely not stand without DLSS 3.

There was some rumblings regarding the idea of motion reconstruction going into gaming, the form it's coming in with DLSS 3 is not what I expected.

DLSS 3 does address an issue by somewhat uncoupling gains to some extent from the rest of the system. Especially with GPUs in some of the higher classes a problem is that that 2x faster GPU in the theoretical sense cannot actually achieve 2x real performance gains due to limitations from the rest of the pipeline. Even 4k with RT (since RT itself is also showing to be heavy on the rest of the system) is not always even fully GPU bound. This situation might get even worse (in a sense) as we transition more towards the current consoles as a baseline (especially if we fully ditch last gen).

We'll have see when the whitepaper comes out for more information if there is anything more on how they are dealing with memory bandwidth. Cache itself has it's limitation (which we see with RNDA 2). It could be that Ada also might in some scenarios be rather limited due to the lack of raw bandwidth growth this generation.

This feels like Turing/2xxx for me personally anyways. The new product stack isn't very appealing due to the hardware configuration and pricing, while the new feature set makes the outgoing generation unappealing as well at this stage.

Hardware aside some of the presentation regarding their Omniverse platform as it pertains to game development (and modding) was very interesting.
 
Last edited:
1. RE8, ACV, TD2 do not support DLSS.
2. All of these are running at >100 fps on 3090 in 4K at max settings. I won't be surprised if they are CPU limited on 4090 in the same scenario - and that this slide shows what DLSS3 brings to such cases. MSFS is 100% CPU limited as can be seen from lack of scaling between all three cards.

Have we all already forgotten that we should wait for proper benchmarks?

Indeed, which begs the question..
3sRPsa2.jpg
 
Very crude raster efficiency gains excluding DLSS using the 3 left games, TDPs and 3090 perf + 350W TDP because 3090Ti's 450W TDP is far less representative (1.3x TDP vs 3090 for 1.1x perf). We can compare overclocked Ti versions when both gens have them:
4070 4080 12GB is 285W, about 1.1x 3090 perf at about 0.8x TDP, roughly 1.35x perf/w
4080 16GB about 15% perf/TDP increase vs 12GB, again 1.35x
4090 about 1.75x perf with 1.3x TDP vs 3090, 1.35x perf/w

Yes there're a lot of caveats, excluding things it's really good at (RT + DLSS 3) is a big handicap, it's marketing slides, limited games, TDP =/= actual power etc etc but even so 1.35-1.4x raster perf/w seems underwhelming going from SS 8nm to TSMC 4N with a big arch change. Isn't RDNA2 around 10-15% more efficient roughly than Ampere in raster? RDNA3 could have a very significant lead in that area if their >1.5x perf/w claims for RDNA3 are good, 1.2-1.3x (more?). DLSS 3 is exciting, RT improvement tentatively looks great, prices aren't fun, wait for benchmarks
 
No we didn't. Titans were a separate lineup, and we were getting "xx80Tis" in GeForce with essentially the same (sometimes higher) performance at 1/2 or 1/3rd of price. This isn't the case with 3090 cards.
Yes, we did. Titans occupied the same niche as the 90's cards do now. The original Titan was marketed as a prosumer card with much higher double-precision performance than its contemporaries until this was dropped and they became mostly gaming cards with huge VRAM pools and a much higher price tag. Titans were also named GeForce cards before Pascal.

Thought this was obvious?
 
What is that even supposed to prove? That the 90's cards have an absolutely abysmal price/performance ratio even compared to the Titans?
Have you read the start of this line of arguing?
"The outgoing king's" performance is different this time because it's 3090 instead of a $3000 Titan and a $1000 3080Ti with the same performance.
 
Have you read the start of this line of arguing?
"The outgoing king's" performance is different this time because it's 3090 instead of a $3000 Titan and a $1000 3080Ti with the same performance.
You might wanna rewrite your sentences because I'm trying to decipher what you're trying to communicate and can't do it.
 
Neat tech, but the tech and pricing feel at odds to me for an era of games where the big budget content are made for cross-gen console with hit-or-miss ports to PC. I'd feel super weird dropping that kind of money on a card and not having a single title that I would care enough to download, install, look at for 3 minutes, and uninstall again. The 3000 series is probably plenty for most people that need to drive their back catalog of games at 4K. I doubt Starfield and Diablo4 next year are going to stress these new cards, and I'm not sure when the first big UE5 titles are expected...? Perhaps I'm just aging out of video games.

So... can this optical flow engine reproject a lengthy sequence of frames a la Oculus's spacewarp to mask huge stutters, or is it merely for producing interleaved frames? What will it look like when it slams into a 1-2 second long shader compilation?
 
The new RT Cores also include a new Opacity Micromap (OMM) Engine and a new Displaced Micro-Mesh (DMM) Engine. The OMM Engine enables much faster ray tracing of alpha-tested textures often used for foliage, particles, and fences. The DMM Engine delivers up to 10X faster Bounding Volume Hierarchy (BVH) build time with up to 20X less BVH storage space, enabling real-time ray tracing of geometrically complex scenes.

From the following link, it seems like taking advantage of the OMM and DMM engines will require using special Nvidia SDKs for Nvidia Micro Mesh's. So the BVH improvements and the alpha-testing improvements may not be realized by most games. Not sure how widespread the adoption of these technologies will be, but I'm guessing adoption won't be good. These technologies are cross-platform and open source.


NVIDIA Micro-Mesh is a graphics primitive built from the ground up for real-time path tracing. Displaced Micro-Mesh and Opacity Micro-Map SDKs give developers the tools and sample code for the creation, compression, manipulation, and rendering of micro-meshes. From fossils to crawling creatures to nature, you can express these assets in their full richness.
 
Again, the naming, why is it so hard for marketing in so many cases?
"4080 16GB" should've been "4080Ti".

Nah, they need to save the 4080Ti name to fill that enormous gulf between the 4080 and the 4090. Based on the gap between these GPU's, the 4080 12GB is clearly a 4070.

Have you read the start of this line of arguing?
"The outgoing king's" performance is different this time because it's 3090 instead of a $3000 Titan and a $1000 3080Ti with the same performance.

The comparison points I was referring to were GTX 970 = 780Ti, GTX 1070 = 980Ti, RTX 2070 = 1080Ti (granted Turing didn't quite hit this one, but then Turing was considered a performance disappointment) and 3070 = 2080Ti.

The difference here is that the price from the GTX 970 to the RTX 3070 went up by only about 52% over the space of 3 generations.

The 4070 has increased in price over the 3070 by 80% in a single generation. No wonder Nvidia wants to convince people that it's really a 4080, and even then it's a pretty rubbish deal given that it's more expensive that previous generations x080 class products and is barely faster than the 3080Ti.
 
Status
Not open for further replies.
Back
Top