Nvidia real-time raytracing

[Nejc Kilar]

Yeah, I was not referring to project Lavina. What I meant was this video ->

As for the NVLink, it sure does seem like we'll have that on all the RTX GPUs. Including the 2070. Should be an interesting presentation on Monday.

Still, I would prefer if AMD had something ready as well. I really dislike the current high end monopoly. It reminds me of Apple and how they kind of shafted their own content creators...

[soso]

Yeah, I was not referring to project Lavina. What I meant was this video ->

I don't think they using hybrid GPU rendering in thad video (GPU+RTcore+Tensor core). Maybe It's just using RT core only. I'm not sure. And I don't know how effective this experimental V-Ray GPU build with a pre-release Quadro RTX 6000 and driver. If you compare it to Unreal RTX video, they are using hybrid rendering (GPU+RTcore+Tensor core). I'm not sure too if maybe unreal engine build itself in the RTX video, is better than 3ds Max & vray for realtime ray tracing capability.

Well just wait the RTX news from other GPU renderer too tho.

[Juraj]

Was that video supposed to be impressive ? Maybe it is, I didn't test it with 1080ti but it's simple non-GI empty scene at 1000px...

I really need to know if it speeds up the Optix Denoiser. If the RTX engine or the tensor cores inside 2080/2080Ti somehow magically alters the performance, that could be dealbreaker.

Otherwise they can go f*** with these prices.

[danio1011]

Was that video supposed to be impressive ? Maybe it is, I didn't test it with 1080ti but it's simple non-GI empty scene at 1000px...

I really need to know if it speeds up the Optix Denoiser. If the RTX engine or the tensor cores inside 2080/2080Ti somehow magically alters the performance, that could be dealbreaker.

Otherwise they can go f*** with these prices.

I know it's just proof of concept but I'm surprised they would show such a simple scene as the first public test of RTX on RT.  People have been seeing 'single object' scenes on VRay RT for years.  I can't really tell what's going on with VRay for Max these days, it has made me hold on upgrading to Next even though I find it (GPU, adaptive dome, auto exposure, etc) intriguing.

[Nejc Kilar]

- Bare in mind people that apparently the RTX 2070 will not be getting NVLink support so there is some product segmentation within the RTX line itself.

(https://videocardz.com/newz/nvidia-geforce-rtx-2070-does-not-have-nvlink) & (https://www.guru3d.com/news-story/nvidia-announces-geforce-rtx-2080-and-2080-ti.html)

- Also, NVLink for the Geforce brand appears to be limited to a speed of 50gbps instead of 100gbps. That is apparently a quadro vs geforce segmentation issue. I don't have any sources to share but I think its listed on their website (nvidia).

- Another thing to note... Some people are saying that the Geforce cards will not have NVLink support for a card by card setup. That means that you will not be able to link more than two GPUs together in a single case because those two will use up four slots (I think it was mentioned there will be  a 3 slot nvlink bridge or something like that). That is unless you are running a fancy setup with PCIE risers I guess. I don't have a source to share but apparently it "makes sense" because the FE cards are not blower style cards anymore and so it is not wise to stack them together. It goes without saying that the Quadro line will be able to have smaller NVLink bridges so there is some segmentation there too I guess.

So all in all... Its weird :))

[Juraj]

The GeForce NVLink still seems fast enough. After all the full NVLink is made to be sufficient for the top range Teslas and Quadros, some of which feature 6k+ Cuda cores, so if the GeForce has half the bandwidth, that is still more than enough for the kind of performance 2080ti might yield.

But does it have all the features in functionality ? Is the memory stacked ?

22GB of Vram is A LOT for GPU rendering.

[burnin]

Have no doubt - "If it's not clearly written, then it doesn't exist."
(Either contracts, bills, invoices, inquiries, news, legislation, obituaries, articles, theses, ... - it's common knowledge & every person should be aware of that.)

So,
a.) for top tier gaming cards (2080/Ti) specs states: NVLink (SLI ready)
b.) while on the other hand, for whole RTX Quadro line it clearly states: "NVIDIA NVLink® to combine two GPUs with a high-speed link to scale memory capacity up to 96GB and drive higher performance with up to 100GB/s of data transfer."

Go for a long walk and drink water. ;)

[Juraj]

:- D

[soso]

Here's the updates from Vlado about the new GPU ray-tracing:
https://www.chaosgroup.com/blog/what-does-the-new-nvidia-rtx-hardware-mean-for-ray-tracing-gpu-rendering-v-ray

[Juraj]

So Vlado hints that NVLink is identical in terms of memory pooling for all RTX cards.

This is super revolution for GPU rendering ?! Christ, I can't even imagine what some people with 4x2080ti will do.

Maybe time to try Vray GPU once they implement support for RT cores. Ultra-speed, support for Corona materials plus all the nice things we are waiting forever for ( VR SCANS ).

[bluebox]

Hey there guys. We're on the crossroads here. Soon we will have to get some more raw rendering power and will have to decide either going CPU or as it seems viable finally (?) GPU. Is my understanding of the topic and the things you guys suggest correct - that I am able to stack 4x 2080ti in one machine and get 44GB VRAM ? As far as I know 44GB Vram would be enough to fit (at least with Fstorm and it's inbuilt texture resizing) virtually any scene no ?

[Nejc Kilar]

Hey there guys. We're on the crossroads here. Soon we will have to get some more raw rendering power and will have to decide either going CPU or as it seems viable finally (?) GPU. Is my understanding of the topic and the things you guys suggest correct - that I am able to stack 4x 2080ti in one machine and get 44GB VRAM ? As far as I know 44GB Vram would be enough to fit (at least with Fstorm and it's inbuilt texture resizing) virtually any scene no ?

Technically yes, that appears to be true. We still need to wait for the official response from Nvidia whether the Geforce NVLink = Quadro NVLink or if it is just a cutdown SLI version of some sorts. Considering that it is called NVLink, it shares the same physical connector and is known to have sufficient speed (50gbps compared to 100gbps on Quadros) I think that it is pretty much the same thing.

That being said, we do know that Nvidia currently only sells 3 and 4 slot NVLink bridges for the Geforce cards. That means you can probably NVLink a bunch of 2080ti although you'll need to use PCI-Express risers in order to fit them on your motherboard.

I have 6 PCI-E slots ready but with the Geforce NVLink bridges I could probably only fit 2 cards in there in a standard chassis without any risers. Technically though, I could easily fit 4 dual slot cards without NVLink (4 x 1080 for example).

Quadros come with 2 slot NVLink bridges so with those you can actually easily NVLink 2 + 2 cards in your system and make it all fit much like with the current gen GPUs.

Hope that helps :) I am looking into it myself too but I will wait till the cards are actually release to see if everything is "as it should be" for rendering :)

[Juraj]

Do you actually see multiple cards being connected ? I only see two cards, even in Quadro/Tesla range.

The 3-4 slot are only to accomodate wider spread between pci-e slots not to cover 3-4 cards, at least that is my understanding.

So with RTX 2080ti gen, 22GB Vram would be the current limit. Which is heck of a lot in gpu-rendering.

[Nejc Kilar]

My bad, I totally didn't address that properly.

You can only connect two cards (hence 2+2) which obviously gives you access to the combined VRAM of the two cards. So if you have 2x 2080 your VRAM total will be 16GB. That being said you can connect 2 + 2 + 2 ... GPUs together. The combined VRAM total will still be the same but your rendering performance will be obviously higher. :)

So yeah, you are totally right. The 3 & 4 slots are there to create some room between the cards. Thats pretty much it afaik.

[Juraj]

I really can't wait for Vlado to post his benchmarks once NDA falls out. His words: You will be positively surprised :- ).

And that is yet without the RTX cores I believe.