Daz 3D is part of 
Connect
DAZ Productions, Inc.
7533 S Center View Ct #4664
West Jordan, UT 84084
Licensing Agreement | Terms of Service | Privacy Policy | EULA
© 2025 Daz Productions Inc. All Rights Reserved.
Comments
I'm still relatively new to DAZ (started last year, some time after my last GPU upgrade to a GTX 1080). Am I to understand that buying one of these new cards means I can't use anything but CPU rendering in DAZ until new drivers are released, which could take months?
The other big rumor with the new NVLInk is that you might be able to pool GPU's themselves.. So if the rumor is true can you imagine having two GPU's working together rendering and with combined Vram well that would be awesome..
Maybe. There was a long wait for Nvidia to get a version of Iray that would work with the new cards to DAZ. It did not take long for DAZ to get it released once they had it. The wait was all Nvidia.
I agree, that presentation was about spin; my opinion of course.
...as I expected, all the talk of a 16 GB 1180/2080 was nothing but smoke. The 2080 will still have 8 GB and Ti version, 11 GB. The one major change though, they support 2 way NVLInk, so if you buy a second card and plop about another 900$ down for the NVLink widgets apparently you will be able to pool memory to get 16GB and 22GB total respectively. So using two 2080s along with the two way NVLink setup will set you back 200$ more than a Quadro RTX5000 which also has 384 Tensor cores, 6 GRays/second ray tracing, as well as also supporting NVLink to offer 32 GB of VRAM with two cards.
No mention of Tesnor cores, which it looks like they are reserving for only the Titan and pro grade lines.
...oh, and also no mention of AI denoising support.
You missed the tensor cores mentioned in the MSI info on Tom's Hardware:
https://www.tomshardware.com/news/msi-geforce-rtx-2080-ti-specs-pics,37629.html
...OK just saw that buried in the text. So the RTX 2080 has the same Tensor count as a Quadro RTX 5000 and the Ti has the same count as the Quadro RTX6000/8000. Didn't see that on PNY's specs at all and you'd think they'd hype that as well, particularly the prices they plan to charge.
Think I'll just sit tight for the "official" word from Nvidia..
Here is a video of the new cards and some very interesting details already known and some not..
Already known:
2080 8Gig GDDR6
2080Ti 11Gig GDDR6
Yeah if they do that to NVLink on the consumer cards then yes what is the point of NVLink but on the other side then people would not buy the Quadro cards, so there would have to be something that really does differentiate between the types of cards.. With that the one thing with Quadro cards is that the hardware as I can gather is specifically designed to take the stress that things like 3D rendering would place on it, unlike the consumer cards which I can gather are not really designed for heavy 3D rendering work loads..
Either way we will see very soon what the new consumer cards will have and what they can do..
With the Pascal cards, the delay was in Nvidia releasing the Iray SDK for the new architecture. DAZ was commendably fast in incorporating this into Studio once this was done.
Cheers,
Alex.
outrider42, since you seem somewhat (totally?) convinced that tensor cores are the major player in any performance improvements we can expect from the new RTX platform, I'm hoping you can put some actual numbers to it.
For example, if I have a Studio/Iray scene that renders in, say, 10 minutes with my GTX-1080ti, and I pull out the 1080ti and put in a 2080ti (or any other RTX card), how long with the scene take with the RTX, and what percentage of that performance gain will be due strictly to the presence of the tensor cores (and not the Optix software, RT cores, SM cores, NGX software, Physx, etc.) ?
The specs I saw for the power requirement for the 2080ti are 285 watts, which isn't a whole lot more than the 250 watts of the existing 1080ti (ie, only a 14% increase). Why would that require water cooling for multiple cards?
Also you seem to imply that more cores is better performance. I'm not sure that's the case. In fact, the chart below shows there is very little, if any, direct coorelation between core count and Studio/Iray performance.
To be clear, I think there may be some confusion about the benefits of NVLink. I don't claim to be very knowledgeable on the subject, but based on what NVIDIA has said it appears to be primarily a way to scale (add) memory of two GPU's, and also provides a direct link (similar to an SLI bridge connector) which effectively bypasses the slower PCIe connection between the two GPU's.
With existing configurations with, say, two 1080ti GPU's connecting via PCIe and doing Studio/Iray rendering, both GPU's are already being fully utilized (all cores) if possible. I don't see how an NVLink connector would necessarily improve utilization of the two GPU's, since you are already effectively adding their cores together.
Presently I think the use of the PCIe bus is primarily for loading the scene into the GPU, and once that's done the GPU's go off and render on their own. So I'm not sure how an NVLink connector would help speed up renders, unless there's now some new inter-GPU communication over NVLink that is involved with render calculations?
If the memory scaling indeed results in an addition of both VRAM capacities that would be great, and I'm hoping that actually results in two GPU's VRAMs adding together when doing Studio/Iray rendering. I suppose we'll see if it's that simple.
It is all interesting, but until we get IRAY updated - pointless; for me a very expensive paperweight.
Another thing to consider regarding the NVLink memory "scaling"...
Keep in mind that fitting a scene into VRAM also requires you fit it into system RAM. My ballpark is that somewhere around 3X VRAM is required in system RAM to hold a scene. I've got scenes that exceed 30GB of system RAM that just barely fill my 1080ti VRAM (10+GB). So if you're going to have two 1080ti's with memory scaling, that means you'll need around 64GB of system RAM to fill the 20GB of combined GPU NVLink VRAM. So if you buy two NVLink cards ($2,000?) you'll also need to upgrade to at least 64GB of system RAM ($600+)?
The devil is, as always, in the details. And yes, the paperweight is looking VERY expensive.
I wouldn't get too excited about tensor cores and RTX, the storm troopers looked incredible in real time, but did you ever wonder why they chose storm troopers and not real people with skin and hair? Becuase just like the AI noise thingy we just got in Daz, it works great on shiny surfaces, but kills all renders that have hair and skin; You loose the details. Tensor cores, no matter how you slice it and dice it have to take short cuts and make guesses, to cut the render times.
True, but technically it's not even the tensor cores that slice and dice and take short cuts and make guesses to cut render times. It's the denoising (software) algorithms that do the real work. And denoising has been around for a long time, so it's nothing new. Last year it was available with Blender, and it sounds like games have been using it for a while too. And it has nothing to do with tensor cores. Which is why I'm so puzzled about the focus on the tensor cores, which are merely some hardware chips that do the specific tasks associated with denoising and AI.
I think the software guys came up with new denoising algorithms a long time ago, and figured out the specific steps you'd need to take to implement them. And then they asked the hardware guys "Hey, I need to do these math steps to do my new denoising algorithm, can you guys make some hardware chips on the next GPU that do those specific steps real efficiently?" And the specific steps were to calculate what are called "tensors", which involve calculations of 4x4 matrices. So now the tensor cores do those specific calculations that the software tells them to. But that certainly doesn't mean that without tensor cores you can't do denoising. It just makes it faster and more efficient, because they're designed specifically to do the denoising calculations.
That being said, denoising can be extremely useful in certain situations, and it can make a huge difference. You can see tons of Blender videos showing the effects of using the denoising feature, even before tensor cores became available. And I'm sure the same applies to existing game engines. But that doesn't tell us anything about how/if the tensor cores will have a significant impact on render times above existing non-tensor denoising algorithms. And it doesn't tell us if, as you say, denoising will be acceptable in terms of quality for our particular renders.
While I sympathise with your frustration, we don't know why DS doesn't have motion blur - there's certainly been noindication that I am aware of that it's simply considered unneeded, so this comment is not accurate.
I did some more research on this denoising thing, and it's pretty interesting...
Turns out the real star of the show is, as I suspected, the software. Optix 5.0 is the latest rendering API released last year, and it includes the denoising engine. Apparently what the engine does is something like this:
Just like you and I can look at an image and quickly determine "hey, that car image is too noisy", NVIDIA engineers came up with a way to generate rules that a computer can use to make the same determination. They took tens of thousands of images and many different scenes, and, based on those, developed rules on what is good and what is bad. That's where the artificial intelligence comes in. You teach the computer to recognize good and bad based on rules generated from a bunch of data. It sounds like they took a bunch of images generated at different render stages from many scenes to determine rules about what noise looks like, as well as how to make filters to clean it up. And apparently those rules are included in Optix 5.x, so when you develop your renderer with denoising it uses those rules to determine what is good and bad, and then filters/re-constructs the image to make it better. And I think the details of how they did that are the real brilliance here.
It's a bit like how you can get an app on your smartphone that, when you take a foto of a flower it will recognize it and tell you what type of flower it is, based on rules generated by a bunch of software developers. Like if you had an image of a flower, and it was mostly bright yellow, and you could detect edges that form petals, and you could count the petals, then you could say "hey, this is a daffodil, and it should look like this, not with all those noise speckles", and replace parts of the image with what your database/rules tell you a daffodil looks like.
So instead of doing time consuming ray tracing to get a final image, you can do some ray tracing along with progressive denoising to guess what the image should look like based on it's knowledge of all those tens of thousands of images. And I'm sure a bunch of other stuff gets included in the rules, like scene data (normals, depth, brightness, surface color, etc.).
Denoising is great for distance, and non-organic; organic can often not look good until the needed rendering has been done.
...last I saw on the Nvida site, a single NVLInk connector was priced at something like 445$ and two are needed for the two way connection. That is 890$ alone, so at 1,000$ per card (based on PNY's projected price) you are talking 2,890$. Yeah, less than half the 6,300$ price of a single RTX 6000 for 2 less GB of VRAM, but still out of reach for many people here.
WHAT??? Wow I missed that. Weren't the SLI connectors pretty cheap? Or am I mis-remembering again?
More things to go wrong with multiple items, but also; if you lose a card, then it's only one card; rendering is slower but still decent.
I'm excited. Time I got a new card; I started saving up when the 1080 was released, and well just carried on saving. :) Better nor me, not so much for Nvidia. :)
...no, SLI bridges (2, 3, and 4 slot) all are priced at 39.99$. NVLink ones as I saw on their site (for the Quadro GV100), are actually 599.99$ each (and again you need 2).
I will wait for all of the RTX buyers second hand Titans and 1080ti's ;)
You cannot compare core counts across generations. Every generation the cores are made more efficient. I have covered this as well, I wrote a post on the CUDA compute of every single GTX in existence at the time. When you look at your file in Daz, it shows your GPU's CUDA compute rating. Every generation this goes up, and sometimes you will see cards released late in a generation have small increases in CUDA compute. The best example of how different CUDA is across generations would be the 680 and 1050ti. The 680 has exactly twice the CUDA count as the 1050ti. But the 1050ti is actually faster for rendering (and gaming.) That is because the 680 released in 2012 and has much slower CUDA cores, its compute ability is 3 generations behind. The 1050ti is also clocked faster, and it is this combination: faster CUDA cores plus faster clock speed that makes the newer 1050ti render faster than the older 680. It is not just one thing. So CUDA counts don't always mean everything, but this was for an older card. Turing will have the -FASTEST- CUDA compute ability yet seen in any GPU. This combined with a high clock and a boat load of CUDA cores will make the 2080 and 2080ti absolute beasts for Iray.
But that's not all. What I just wrote doesn't even include the new features of Turing. Turing is optimized for ray tracing. I don't have all the answers, because this is all speculation. But I have that Titan V benchmark that none of you want to believe is accurate. I don't know why you refuse to believe that benchmark, it is specifically for Iray. We use Iray here. The company behind it has been doing Iray even before Daz3D got Iray. It is a real benchmark, not just something they made up. And nobody else here has a Titan V Iray benchmark to show, because nobody else has benched a Titan V for Iray. If someone would like to show a Titan V benchmark for Iray that directly contradicts the bench I linked, I will happyily shut up and leave you lot alone. Until then, we do have a benchmark for IRAY that shows a 2.5X performance gain over Titan Pascal and 1080ti using a Titan V, and not a single one of you has proof otherwise. Also this benchmark does not use any denoising. So this is not cheating.
The Titan V is a beast, but in other typical applications it is only 30-50% faster at them. So the question would be, exactly how is the Titan V so much faster specifically at Iray? CUDA by itself has not made that huge a leap in one generation. So, tell me what the Titan V has that previous generations don't. 640 Tensor cores. By the simple process of observation, one has to look at how this speed is achieved and conclude that Tensor is making that difference. Vray, Octane, ect, I bet you none of those make proper use of Tensor at all, and that is why they don't see the same boost with a Titan V that Iray does. After all, Nvidia makes Iray, I cannot stress this enough. It should be a no brainer that Nvidia would update Iray to use everything the Titan V has, including Tensor. Iray is a product they sell, you might not pay for it, but Daz and anybody else who licenses it does. By giving Iray an advantage, this could be what gets Iray to become more mainstream. Nvidia stands to gain a lot of things with this.
But there is something that Turing has that even the Titan V lacks. What is it??? RTX.
So, in summary...you have over 4300 of the fastest CUDA cores ever made, combined with 576 Tensor cores, good clockspeed, combined with RTX to give the 6X performance gain claimed. You only see this performance gain in software designed to use ALL of these features. It just happens that Iray is such a software.
Nvidia has been looking to ditch SLI for years. Fewer and fewer games support it, and even then there are frequent problems. That is where NVLink comes in, and I have wrote about it, too. The idea behind NVLink is to make the PC see the linked GPUs as one single GPU. It combines all the GPU assets together. I am not making that up, that is precisely how it works. It does this because of how fast the connector is. SLI was just not fast enough. PCI is also not fast enough, not even close, and that is why when you render currently, the entire scene must load on to every GPU. The connection in NVLink is as fast the VRAM in some cards...that is how it is able to do this, it is as fast as VRAM! Think of it as manly Gurren Lagann combining. A big version of NVLink is what powers the monster DGX machines that Nvidia has been making for a couple years, and this bigger NVLink is the secret to how Nvidia is able to make such massive VRAM stacks in these machines.
I am assuming that consumer NVLink will cost far less than the Quadro links for sale, after all, those links cost almost as much as a GTX GPU itself. Everything for Quadro is super expensive, so it makes sense that consumer versions will cost a lot less. But they will still cost a lot more than SLI ever did. That is why I suggest $200. The big question is if Nvidia does something to nerf the consumer version, and limit the ability to pool VRAM like its bigger version. Obviously a PC needs RAM, but I would assume somebody buying two 2080's or 2080ti's plus NVLink would be using a pretty large amount of system RAM, or have sense to buy enough. If NVLink does pool VRAM, I plan on building a new machine. Well, I plan on that anyway, but it would be further up my to do list.
NVIDIA tells us what's what at 12 Noon Eastern on Twitch. https://www.twitch.tv/NVIDIA
Ha, can't believe I'm falling for this again but I'm slightly hyped
Well, can be pre-ordered at Scan
https://www.scan.co.uk/shop/computer-hardware/gpu-nvidia/nvidia-geforce-rtx-2080-ti-graphics-cards
£1050 - £1333
Pre-orders direct from nVidia: https://www.nvidia.com/en-us/geforce/20-series/?nvid=nv-int-gcxxpn-56591
$599 for the 2070, $799 for the 2080, and $1199 for the 2080Ti
NVLink bridges for them are listing at $79. Not sure if they are actual full NVLink or not, but that's what they have on their naming.