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
© 2026 Daz Productions Inc. All Rights Reserved.
Comments
Please educate me. I currently have two 780Ti, of which on paper contain more cuda cores than two 1080s. BUT, because of the 1080 clockspeeds there should be a huge difference in performance? Correct?
Thanks.
Both the clock speed (which tells you how fast those cores will finish stuff) and the KIND of cores. The Pascal cores should be somewhat faster than Maxwell architecture on a clock-for-clock basis (but it's not a lot). But combined with higher core/memory clock speed, it adds up pretty fast.
The 780Ti comes with 2880 cuda cores. That's actually more than a standard 980Ti (which has 2816, iirc.) But it's stock core clock would be 875 MHz. That's just slightly above HALF the speed of the stock 1080 (1600MHz).
To give you an approximate performance ratio:
780Ti : 2880 Cores * 875 MHz clocks = 2520 GCoresClocks
1080 : 2560 Cores * 1600 MHz clocks = 4096 GCoresClocks
So basically, the 780Ti can perform 2520 billion CUDA operations per second. The 1080 can perform 4096 billion CUDA operations per second. So about 1.63 TIMES faster than the 780Ti, NOT including architectural improvements, and the faster memory speed as well.
So TWO 780Ti cards would be about 20% faster than a SINGLE 1080 on a cores/clocks basis. Each 780Ti consumes about 250W, where the 1080 consumes 180W. Add in the architectural benefits from Pascal compared to Kepler (which is the architecture the 780Ti chips use), I'd expect the 1080 to be right on par with dual 780Ti cards....maybe faster in some cases. This is assuming stock speeds on both cards, of course. And the 1080 consumes 180W, where the two 780Ti will consume a whopping 500W.
Yeah, I saw the rumours too but I also read that the consumer cards (1070, 1080) would not be available until late 2016 or early 2017. In my mind, it was a case of use IRay in CPU mode for another 9 months or bite the bullet and buy. Clearly - in hindsight - the long delay rumours were not to be believed. I'm kicking myself now, of course.
Have i missed something here? i was always under the impression that iRay and other Nvidia based cuda renderers (octane) got their speed primarily from the cuda cores, and that clock speed on the GPU had very little to do with it. I was told that a gpu with 1500 cuda cores will be pretty much spot on half as fast as one with 3000 cuda cores.
If that is wrong, can anyone point me to do some information that proves that wrong and explains it?
if you are beleiving that a 1080 is twice as fast as a 980 because Nvidia said as much in the reveal, that claim would have been said using gaming as the basis for it, because that is primarily what these cards are made for. For rendering, can be very different
Right now I'm very happy with my new 980TI. It's taken big complex renders down from days to hours, and simpler scenes from hours down to minutes. Of course I managed to blow the memory limit, but even if I had a Titan X, I'm sure that would still have happened -- we pile stuff on until something gives!
I've never been a fan of the bleeding edge when it comes to hardware -- too many bugs tend to get found. I'm quite happy to wait until the first few batches are sold out and things have stabilised. By then the reports will be in of how good it really is for Iray. (and I might have paid off my current kit!)
marble: Well, given how slow bounce light works... Iray, even in CPU mode, can be competitive for realistic stuff.
Sorry, Will - I'll take some convincing of that. Unless you have a different definition of competitive to mine. Nevertheless, the GTX970 is performing as well as I had hoped so it will have to last until I can beg, borrow or (otherwise), enough cash to upgrade. By then, there might be something wowing us all into wondering how we ever put up with those sluggish Pascal cards :)
I had a complex scene. Rendering it at 2160x2160 with a bounce light looked like it was heading for 3-4 hour renders.
I then switched to Iray and got to 99% convergence with quality 2 or 3 (I think?), same size, it took me somewhere between 1 to 2 hours.
And that's with AoA lights in the first case and Advanced Skin, which are about as fast as you can get in 3DL.
As for non-realism, I can render fairly complex scenes in Iray pretty quickly, provided I avoid SubD and SSS. Whether it's faster or slower than 3DL depends wildly on a number of factors, though cool cartoon-shaded 3DL can be pretty dang fast.
Enjoy the GTX970, Marble. You should be fine with that for some time. We can't all be on the bleeding edge. My GTX780ti does just fine for me and it's better to wait until everything shakes out. We don't know if Nvidia has optimized the new cards for CUDA rendering yet. They obviously have for gaming frame rates, VR, power consumption, etc. I remember reading all the frustration on the Octane Render forum when some new cards came out a couple years back. They were great for gaming but showed they hadn't been optimized for CUDA rendering and were slower. If memory serves that's why Cath (MEC4D) waited and jumped from the 700 series to Titans.
I have a titan x, and i was looking to buy another, now i'm not so sure. From what i've read the 1080 has higher clock speed but less cuda cores? I understand gamers will be all over it but is it such a huge leap for GPU rendering, specifically IRAY?
That is a 'yes and no' thing...it is very true with CUDA cores in the SAME generation...ie: clock speed isn't all that important if you are comparing two Maxwell cards, but with extreme differences, it can be a factor (like with a high OC vs a stock card). It does become important when crossing generations...so comparing Maxwell to Pascal, clock speed does become a factor. That's pretty much always been the case, though...
This is really not the case. It's a bit more complex.....
The CORE clock, like with a CPU, is how many cycles per second the core executes instructions. Within the same architecture, it is a linear improvement on speed.
Between architectures, some (or all) instructions on a core may be improved to require fewer cycles to execute. This means that between architectures, some instructions will execute faster. Some may not. Some might actually increase in execution time, in order to allow for other benefits/advantages. But in general, a newer core architecture will run its instructions faster than the prior architecture.
CUDA cores run a program on the GPU. All the CUDA cores will effectively be running the same program, but with different data (in parallel processing lingo, we call this SIMD...Single Instruction Multiple Data) just like the vector processor inside most CPUs (the SSE/3DNow/etc vector instructions.)
If the clock speed is increased, the execution time for a given instruction goes down. However, the distance the electricity has to travel through the pipeline doesn't (for a given processor.) This means, that if you OC, there is a point at which you'll start getting errors, as the pipeline simply can't move the electricity fast enough through the gates. Smaller gates (i.e., smaller fabrication process) means shorter distances, and higher possible speeds.
So, even between architectural changes, CUDA is VERY core clock dependant. The same cards, one OC'd to TWICE the core speed (assuming it was stable) will run TWICE as fast. (this does assume the memory can keep up, which isn't always the case.)
Generally, MOST cards run at a pretty standard stock clock speed. OCing is generally fairly small amounts (only 5% - 15% maximum) due to the increase in heat produced (higher speed means more current to switch faster) and the card overheats quickly beyond that. This is why most current OCing beyond about 10% requires water or cyro cooling to remain stable, and requires much larger power supplies.
The differences in OC's from the factory demonstrate this. Typical OCs are usually in the above ranges. So the differences only account for small gains (like 5% - 15%) so it's not a huge difference. If the stock clock is 1000MHz, you'll see cards with OC's of 1100MHZ, maybe 1200MHz in special cases with extra cooling. So performance wise, the clock doesn't LOOK like it matters. But it DOES, when the clock speed increases get significant.
For gamers, these are more noticeable.....if they are running at 1000MHz and getting 60 fps, then OC to 1200MHz, they're likely to see 72 fps results. To a gamer, that's huge. With IRAY, what you will see is iterations completing faster. If the render takes 30 minutes at 1000MHz, you'd see it go to maybe 25 minutes. (24 minutes, IF the time is JUST the rendering loop not counting setup/load time) If the render is taking 5 minutes, you'll barely notice that kind of increase (it'd go from 5 minutes to about 4 minutes.)
But when you have a core clock boost that's significantly higher (as you often do between architectures.....Kepler ran around 875MHz, Maxwell at 1000MHz, about a 15% increase) AND the new architecture increases the instruction speeds, you get a bigger boost. Say the most frequently used operation took 5 cycles on Kepler, but went down to 4 cycles on Maxwell. That's a 20% increase in speed right there. If other instructions got similar boosts, it gets closer to that level of boost across the board. This is why in operation, increasing clock speed is linear improvement within an architecture, but isn't linear between different architectures.
When comparing between a huge jump in clock speed, you'll see it even more.
ALL cpu and gpu work on these basic principles. The whole 'shorter distances = faster processing' is the big reason we keep trying to make chips (and thus the logic gates on them) smaller. Going from a 28nm process to a 16nm process is a massive drop in size. More gates per chip, but more importantly, those gates are closer together. And with the new chips being a Fin-FET process, they run at 1v logic. So lower power consumption and heat as well.
Pascal is a huge boost. Between the lower voltage/power, faster clock speed, and enhanced architectural improvements.....it gets BIG boosts. I studied microprocessor design back when I was an EE student. They aren't blowing smoke.....it isn't all advertising hype.
That said, a lot of it will come down to driver implementation, and how well the CPU can keep that GPU fed and running. For IRAY, this is less of a concern. Iray performance is going to get a big boost, just like any of the CUDA-based software. GAMING performance will still also be dependant on BUS size and speed. The 1000 series didn't boost that like a lot of people were hoping (the 1080 and 1070 are listing 256-bit bus widths.) So we may not see quite that same level of performance they claim, in some situations. If they'd upped the bus width to 384-bit, or higher (say, 512-bit like in the GP100) it'd be just as much a beast for gaming as for rendering.....
It occurs to me that it might be good to explain the differences between gaming and rendering from the GPU perspective, so people will understand WHY these changes matter, as well as how.
Since GPUs went to massively multi-core parallelism, their individual cores (CUDA in nVidia GPUs) are still just a processor. They are more specialized, with instruction sets that are geared toward the rendering pipeline. As GPGPU computing became more of an item, the chip makers started including more general-purpose instructions on the GPU cores. These particular instructions aren't as fast, and are usually more accurate than the other instructions.
When running, the various cores are loaded with a few specific programs. These are the rendering pipeline code, with varied shaders. A given core will (if things are well-optimized) stay with the same 'program' on it as long as possible. Then, the geometry and texture data are fed through that program, generating pixel data to go to the screen. The program on the core just does it as fast as it can.
As the view 'in-game' changes, or as new objects have to be loaded/rendered, or have their shaders change, or new textures loaded.....the driver has to pick cores and load new data. IF the actual rendering algorithm has to change for a new object, then a new program may have to be loaded for some cores to handle those particular polygons. This happens a LOT in gaming. This is why the bus-width (how much data can be sent per clock) and the bus-clock (how fast each can data block can be moved) matter so much. Also, multi-pass rendering will often require the output from one set of cores as inputs to another set.....and those transfers happen via the data bus. Bigger/faster buses means faster moving of data.
For iray or other rendering-on-GPU, this isn't as much the case. The 'program' running on the CUDA (or whatever) cores is pretty much identical. It's more complex, longer, and has a lot more inputs and outputs. It uses higher accuracy too. Once everything is loaded initially, it pretty much just runs on the loaded data. No additional transfers occur except to update the result set off-chip. So bus-width and speed only really affect how long it takes to get the scene onto the GPU. Once it's there, it has little effect.
Iray is basically just a very sophisticated and long 'program' that runs on a CUDA core. It takes very specific data and parameters, and outputs very specific pixel data. And it takes a long time to execute (comparatively) to standard 'gaming' pipeline programs. It gets loaded onto EVERY core, and the data set is fed through those cores as fast as it can be.
So, the number of CUDA cores, their core clock speed, their architectural generation, the bus-width and speed, and the memory size and speed ALL come into play as to how 'fast' a given card is. The driver implementation will have a LOT to due with how well the theoretical performance and the actual performance match up. Older drivers may not even know about more sophisticated features and instructions, which could provide significant speed boosts, so they never utilize them. That is why drivers can cause performance issues. They may not know about a bigger bus width on some cards, and only load the amount of data they know could fit.
Hopefully, this and my last post have helped a few interested people understand how and why GPU numbers translate into theoretical and practical performance. The real nitty-gritty of the actual gate logic, microcode decode and execution blocks, and more get WAY complicated to explain....though it isn't too hard to understand. Simple logic circuits are pretty easy to grasp, and everything is built up from that foundation. The underlying electrical stuff is where it gets a bit hairy, and I've tried to simplify it a bit here for everyone to get a better idea of how and why it works.
Thanks hphoenix. Very informative.
Thanks hppheonix, makes more sense now. Sounds like you have some experience in this field.
So using those calculations you used above, in theory if i have an overclocked version of Titan X (1291 mhz) the theoretical performance will be just a little less than a 1080, but with 4gb more vram. im happy with that
im eager to learn what the pascal equivalent of a titan will be though
Yep. An overclocked Titan-X would be very close to a 1080, with more VRAM. The Titan also has a wider bus, but slower memory.....so on that side, it kind of averages out. The big difference is how MUCH you could overclock, and the power consumption/heat.....the 1080 wins there by a good bit, thanks to the new 16nm FinFET fab process chips and 1v logic.
It will probably be close to the GP100....which is the current iteration of Pascal used for the automotive industry experimental self-driving cars. I'd expect 16GB of HBM2 VRAM, 4096 CUDA cores, and a 512-bit bus width. Probably clocked very similarly to the 1080/1070. With those stats, you'd be looking at roughly twice the performance of a current Titan-X. Probably would have a version available with 32GB of VRAM as well, at a premium price. As for that premium price, I'd expect it to cost more than the current Titan-X. For a consumer version, I'd bet the price would be around $1500. If they go with a dual-chip solution like they did with the Titan-X, you might see even HIGHER performance (potentially 8192 cores and 32GB VRAM) if they use the Pascal architecture.
Though, I'd honestly expect them to hold off on a Pascal-based Titan. They'll try to milk the 1000 line for money first. I'd be betting the next-gen Titan would be the first consumer-available Volta card. Which means at least a year before it'll be out. That's just my own prediction, though. We'll have to wait and see what nVidia does.
Edit: Just so everyone knows.....while the new pascal-based cards are looking pretty spectacular, that by no means indicates that it would be best to run out and buy one. They're brand new tech, the drivers are effectively vaporware at the moment, and they are unproven. The SPECS look great. Let's see how they actually perform in the real world. I expect they'll do well, once they iron out a few initial bugs. And the cost/performance on them is fantastic. But if you just bought a 970 or a 980Ti or a Titan-X, you should hold off on buying just yet, unless you aren't concerned about cost. If you've got money to burn, by all means get one. But if not, you should not kick yourselves too hard for not waiting. Maxwell cards are still VERY good. And given the performance/price ratios of the 1080/1070 series, they will HAVE to drop in price....why would anyone pay the same for an older and less powerful card? The real question is how much they'll drop. And plenty of used cards are going to hit the markets when the bleeding-edge race people want to have their Quad-SLI-1080-monster-my-rig-is-better-than-your-rig-bragging-box that cost as much as a car and gives them so many FPS their monitor has a digital anuerysm......
Think you meant Titan-Z there.
A dual chip 32GB Pascal Titan-Z would be a pretty tasty prospect for current Iray use. Honking power, and still with 16GB working space (effectively, for Iray). We'd soon max out scenes to fill it up, though.
The GTX1080 seems a terrific buy for new builds, upgrades from old GPUs or second cards. I'll stick with my Titan-X for now, and owners of other high-end 900 series cards should probably stick too.
Oh, now I feel like the second-class citizen because I'm thinking my new GTX970 probably doesn't qulaify as "high-end". Should I be aspiring to wealthy nerdhood?
In my view the my GTX 970 is mid-range, better than the budget 100-150$ cards, but still well below the top cards like titan X. It still gives me 10x boost over CPU alone, so I am happy enough. I need to decide whether to get a second GTX 970 at some point if the prices drop significantly, or to get a 1070 instead. My understanding, is that having 2 cards does not double overall performance, so it might be better getting the Pascal later (it would consume a lot less energy than two 970s as well).
At this point, I would definitely hold out to see what performance the 1070 and 1080s bring.
I have dual sli, and it's pretty solid for rendering but for gaming SLI has been pretty miserable lately. It USED to work so well a few years ago. Now, I think we're such a small niche there's not much effort into the SLI drivers.
I'm really curious what they'll announce for Pascal with the Titan series. From the performance numbers they are throwing out, a 1080 is an improvement over Titan-X but not a huge one.
Okay, a little refresher on what is considered 'low-end', what is 'mid-range', and what is 'high-end'......
These were terms coined by the industry. Advertisers don't use 'low-end', the call that market segment either 'budget' or 'saver'....I guess they think that sounds better. And remember these ranges have some 'bleed-over'....they aren't absolutes.
In current USD terms, the 'low-end' is anything that prices in at $150 or lower. Inflation has moved that point over the last 3-4 years, it used to top out at the $125 mark.
'Mid-range' cards now fall in the $150-$300 range.
'High-end' cards (also referred to as 'enthusiast' or 'gaming' cards) start around $300 and go up from there.
The only higher category is 'professional', which is usually reserved for cards that costs in excess of $1000. The Titan series was borderline when it was released, as a kind of bridge between consumer level and professional level.
Naturally, anything priced in the edges of a category is going to be harder to categorize. The GTX 970 is one of these. It's right at the border between mid-range and high-end. The GTX 960 is firmly in the mid-range, and the GTX 980 is clearly in the high-end. The GTX970 is the 'bridge' card that spans the big gap between them. A 950 would be much the same positioning, between the 960 and more 'budget' cards (940 and below.)
So the correct response is the GTX970 is BOTH. It's high-end (for those who have mid-to-low end systems) and mid-range (for those who have mid-to-high-end systems.)
So don't despair. The price/performance ratio of all the 900 series cards places the 970 as the most cost effective card. It's a damn good card, and a damn good deal, even with the whole 3.5GB futz.
I read about the 3.5GB thing. I have tried to compensate by driving my Dell display from the Intel Skylake i7 CPU. Seems to work fine on HDMI and, I'm told, that releases more GPU VRAM for rendering.
According to DAZ_Spooky the slower 0,5 GB of the 970 is not an issue for IRay and DAZ can use the whole 4GB
I have two GTX 980s. I am thinking of upgrading to two GTX 1080s when they become available.
Just to address something earlier in this thread: while it is true that the CUDA core count is a critical number, clock speed is also important in that iray is a calculation-sensitive task. In other words, the TFLOP number is the number we should all be looking at if we're examing architectures that are similar to each other. That is, older CUDA versions were single-precision and adversely affected performance.
To that end, the TFLOP/s number is typically calculated by CUDA cores x 2 FLOP per cycle per core x clock speed... which puts the 1070 slightly below the Titan X.
This makes sense since the Titan X has a *much* larger CUDA core count than the 1070 while the 1070 has a high clock speed.
As for the gaming and VR performance, this is most likely due to a change in the scheduler (if people remember the AtoS DX12 stuff from last year), async compute solutions, and other VR rendering improvements that would directly impact its performance.
Yes, since the main bottleneck for iray is computational and not memory access related...
this looks like the best thread to get answers been planning on getting a new computer and got a decent budget for a pretty good one and basically this is the first and probably last time can get a really good one many say build your own but for various reasons I can't, lot of places that make custom pcs what's on offer mixed bag what parts your allowed and price is a mix, first was thinking of the Dell Alienware 51 with the three gtx 980 cards the Australian version has less stuff than the US one btw also heard lots of bad things about Dell. But then looked into Origon and was at first considering the Genesis with 4 Titans but bit over my budget wanted to spend, so then I was really eyeing off the Chronus Pro with 4 hard drives and two Titans in the midtower most maxed out version of this model except for the version except for the dual Quadro k5000 cards but that doubles the pc price. But then the GTX 1080 came out and well I'm back to eyeing off the Genesis with 64GB memory 2TB SSD and 2 TB MechDrive, it's 5 slot hard drive cage so I can easily add more + the lower bay can add more drives making a total of 35, but what really pushes it for me is now it has the option of 4 of these GTX 1080 cards in SLI which makes this model not much more than the Alienware with oculus and the Chronos so from what I'm gathering from what I've read and watched everywhere and I'm not that much computer tech minded btw is 4 of these GTX 1080's better than 2 Titans making this pc the best option? Like I said this is the first and more than likely last time I can get something like this and I want a really good pc going to get a lot from for a very long time and I want to do renders with a lot of characters and objects at high detail as well as maybe trying zbrush maybe others like marvelous designer and substance painter and others try my hand at making my own daz content at least for me-recreating my creatures and other stuff. But I'm almost 100% sold on getting this Genesis with the 4 GTX 1080s but I'm still interested in hearing if others think these cards are good for what we do makes this pc more affordable but still a lot of money if not as good option as the other pc with the two titans about $2244 difference
Renders access the full 4GB (well games can too, but there is a slight performance hit).
What the 3.5GB refers to is directly accessible by (iirc) the IOPs or ROPs (I forget their exact name); in the 970, the 0.5GB is accessible via R/IOPs that access other memory too.
At the moment the GTX1080 is only theoretically good for Iray. The hardware potential is there for fantastic performance, but we don't know what Nvidia will do with the all important drivers. You should wait for Iray specific benchmarks before commiting to a purchase.
Iray doesn't seem to be as efficient as some other GPU accelerated renderers at utilising multiple graphics cards. A two-card solution seems to deliver the biggest benefit, with additional cards suffering from diminishing returns. Four GTX1080s is probably an expensive overkill. And you'll still only have an 8GB working space.
See this comment (and thread):
http://www.daz3d.com/forums/discussion/comment/1173766/#Comment_1173766
Mec4D is well experienced in Iray, Daz Studio, and suitable hardware.
thanks well like I said I'm not that much knowlegdeable when it comes to computer specs so a lot of this goes over my head but I like to be overkill than underkill and not be able to do and restricted in what I can and can't do with my renders and such I just like to create what I envision in my head and I keep wanting to do bigger and bigger scenes with a lot in them as well as detail. Currently my main comp is my laptop Aspire V17 Nitro with I7 4710Q2 2.5Ghz-3.5, GTX 850M with 4GB dedicated ram, 16GB ddr3 memory, 1000GB HDD. This not long ago suffered the black screen of death couldn't turn it on sent to be repaired came back with everything wiped off only basic windows and most of those functions can no longer use access including camera and windows store. So took me a long time reinstalling daz and other stuff at the time over 100 GB of Daz content and I've bought a LOT MORE since then so I want to get and can get a pretty good desktop like I said may do a little gaming but mostly console player really so it's really our artwork and trying my hand at creating my own content maybe video so I want something I'm not restricted in what I can create and I thought the idea of having 4 of these new gtx 1080's and how everyone's raving about them would be a good option in the Genesis which is a beast of a pc from sounds and looks of it + it's easy upgrade options plugging in additional drives like the old game carts in consoles and not having connect cables + the 64gb memory someone else said go 64 if I could power supply is 1500 this build with 2tbssd+2tbhdd was bit over $12000, the other one The Chronus in the midtower option with 2 Titans did have 4 drives but now has only 3 for some reason picked a 2TB SSD and 2x $TB HDD, 32 GB memory, 1200Watts with some other options comes to just over $9000 this is including warrenties if I bought those though I could get a local repairer for $100. So going for the genesis combo may not be practical then? and the chronus better option? I remember reading sickleyeild was planning on upgrading her 2x gtx 980Ti to 1 Titan as a better option and a Titan has more Cuda Cores which said important and read the 1080's don't have as many as the Titans still more than the 980 but 4 in Sli is a lot more but then from what you and Mech said doesn't mean all of those get used in Iray and are wasted but so two Titans though would still be better with the amount of cores they have and would use more of them probably compared to the 4x 1080's + saving around $3000 better option?
just read Mec's latest comment that she might prefer 2X 1080's over her Titan X as it gives her more rendering power about 40% that's why I thought 4x 1080's would be render powerhouse so many pros and cons on those two builds deciding on