Concerning AI FP8 made reference above in your post
1 Like
Despite of its marketing shenanigans and price pumping, Iâm very intrigued by their racer rtx demo, the demo shows, apart from its near offline rendering level of visual fidelity, the great physics representation, one snippet shows how wonderful the suspension works, the others show how dynamic the cars move and interact with the environment, not that highly realistic car physics is a recent thing, but from what they imply, its all physics based, that rather opens the door for so much in the future, or is it just physx in new incarnation, guess weâll have to wait and see the demo to release in Nov.
2 Likes
Jensen Huang; Chips going down in price is a thing of the past
Is Mooreâs law actually dead this time? Nvidia seems to think so | Ars Technica
It almost sounds like (from this article) that Nvidia will aim to keep the price/performance ratio pretty static from now on. That could mean the stagnation or even the death of GPU-powered rendering, since 500 dollars 5 years from now will only guarantee the ability to render the exact same scenes at the exact same performance as you are doing now with your existing card (which is assuming is a 500-dollar chip).
If this becomes the case for AMD and Intel as well, it will largely mean the end of actual performance upgrades for most people (in favor of just keeping the current machine until it breaks), as an actual upgrade will mean spending at least double what was spent last time.
1 Like
Donât think so, unless there is a cartel even if at foundries level -but there will be a lot of redundancy in founderies with UE and USA subsidies to build new factories in case China attacks Taiwan. Market pressures drive the price down if there are enough competitors. If Intel can establish itself is another player. it would also mean that if Neidia increases performance 2x, the price also doubles. Note that i am not talking about inflation.
DLSS enabled reduces consumption from 461w to 348w temperature reduces frrom 55ÂșC to 50ÂșC.
1 Like
So Nvidia might be generous enough to increase the performance for the price (at least for now). However, it does not change the fact of Lovelace being widely panned by gamers especially as the AIB designs are being unveiled (humongous hulking bricks with, potentially, enough power consumption to strain home circuits and enough heat production to make rooms uninhabitable). Nvidia might end up doing the unthinkable and outdo AMDâs misstep with the Vega architecture.
Sure you have DLSS which is supposed to help, but many see this as just a workaround to justify the price hikes since it may also come with image degradation in cases (ie. who cares about artifacts when you get the frames).
1 Like
It seems at 461w in above article temperature is 55ÂșC it do not seems excessive.
Some Q & A
Does that mean the eevee viewport could possibly get a 4x boost if DLSS is used?
from my understanding DLSS is just an upscaling ai so it takes a lower resolution image and upscales it in real time, so I donât see why there isnât an offline version of it for cycles where you can render at 720p and it upscale it to 1080p or 4k, or just make the viewport run smoother
1 Like
I asked that in Blender Developers forum, but they deleted my post and said i should post in ârequestsâ board⊠i have not made it yet.
In theory should work either with Eevee and Cycles.
Blender does not support vulkan (and dx12) yet and wont in near future.
Eevee does not support motion vectors yet, eevee-next will.
But Eevee developers are focused more on making it actually working on basic level, than adding extra features that are vendor and hardware locked 
.
1 Like
What is the relevance of Vulkan and DX12 in this case?
Oh, also DLSS SDK conflits with GPL.
FSR( and other GPUOpen goodies) is MIT so it is ok to integrate, but still requires Vulcan.
1 Like
Thatâs quite exciting as I intend to make heavy use of Unreal Engine 5!
1 Like
This strikes me as a misleading claim from Nvidia. DLSS 3 is not a pure upscaling technique. It predicts and inserts new frames based on priviously rendered frames.
Iâd extrapolate, that therefore it wouldnât be helpful for single frame renders (and that seems to be the case, because they donât advertise DLSS 3 for offline renderers). I guess the Unreal Viewport could run smoother. But Iâd be hesitant to use DLSS 3 to render image sequences in Unreal due to artifacts caused by the technique.
As far as I remember DLSS 2 was not a general purpose solution and it had to be fine tuned to each specific game.
DLSS 3 uses optical flow additionally, to make it a (more?) general purpose solution. To be honest, it is hard to believe that additionally using the optical flow and previous frame(s?) would not work on RTX 30XX or even RTX 20XX.
I agree with you that DLSS 3 is likely not a good fit for offline rendering. However, the exact same technique is likely quite suitable for offline rendering if the neural network is made larger and slower.
"(âŠ) DLSS 3 will take the existing inputs â frame data, motion vectors, depth buffer, and the previous frame(s) â and adds a new Optical Flow Accelerator. (Itâs actually not entirely new, as Ampere GPUs had a slower OFA, but apparently itâs not sufficient for DLSS 3 â yet.)
(âŠ)
The OFA is a piece of fixed function hardware, dedicated to generation the optical flow field. This is sort of like a motion vector map on steroids. The Ada OFA has a rated performance of 305 teraops, whereas Ampere GPUs had an OFA rated for 126 teraops (integer operations of some form). Again, that suggests a future update to DLSS 3 could enable the algorithm on Ampere GPUs, though perhaps there would be more of a quality loss.
For now, DLSS 3 will only work with RTX 40-series (and later) GPUs (âŠ)"
Source:
There are no mentions of 20xx series in the article, so I assume OFA wasnât included back then and even when introduced in 30xx series, it wasnât fast enough to make it viable, so they didnât talk about it.
I assume it needs several buffers to interpret the image and interpolate. A motion pass, a depth pass⊠just my speculation but I wouldnât be surprised.
Usually, motion vectors coming from a 3d application are way more precise than optical flow maps computed between two images. When dealing with motion vectors, there are a few additional things to consider.
What exactly is on steroids?