Is CUDA number a critical parameter?
Dave McH
Posts: 49
Hello everyone!
Just being wondering if better GPU will decrease rendering time dramatically.
As we know GTX Series have
1050ti- 768 cores
1060- 1280 cores
1070ti- 2432 cores
1080ti- 3584 cores
But will it really help if I go from 1050 to 1070? Yes, it will be better, but what will I get?
I mean if anyone made or saw results of tests, like how long does it take to render the same scene with different GPUs?
It seems to me that double CUDAs won't make rendering twice faster. Am I right?
And another question, to not to create more threads. What lighting is faster to render: light sources, iRay emitter or HDRI map?
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
There is a thread with benchmarking information over in the Commons forum.
Iray Starter Scene: Post Your Benchmarks!
You're right that it isn't a linear scaling. More cores obviously helps, but you'll also run into data bandwidth and other limiting factors.
On your second question, HDRI tends to be faster if you're doing outdoor scenes. However, Iray is notoriously slower when you're dealing mostly with interior scenes which are indirectly lit from outdoors. In those cases, Ghost Lights and other lighting options may help speed your rendering. In general, the more light has to bounce to get to your surfaces, the longer it will take for the render to converge.
The improvement is not strictly linear, but there's enough of a benefit to make it worth your while to get the 10xx-series card with the most CUDA cores you can afford. I mean, why not.
No one can provide a benchmark benchmark because there are too many variables involved. Someone could create a scene that doesn't really test the benefits of the cores, or represent the type of rendering you do.
Light provided from a *good* HDRi map tends to converge the fastest, but it depends on whether the light provides good overall cover. No light is really better if the illumination into all or part of the scene is indirect. Indirect lighting (plus also lots of reflections, complex texturing, and a few other things) consitite the biggest drags on rendering time.
@Dave McH
HDRI faster than Photometric spot faster than Mesh lights (emission). Double the CUDAs will be closer to twice as fast than not. The other parameter is VRAM on the card. The 1050ti will drop out with any render over 4GB, while the 1070ti will render a scene up to 8GB. For a scene over 4GB the 1070ti will be many times faster than the 1050ti (whic is not bing used).
Yeah, CUDA cores are not a direct indicator of Iray performance. For a bunch of reasons. For example, GPU architecture matters. Some GPU's with fewer cores can outperform a GPU with more cores because of advances in architecture, and/or changes in drivers and the "engines" the different GPU's have. NVIDIA designs their GPU's and drivers to perform certain tasks more efficiently than others, and it gathers certain components of the GPU together to make an "engine", which is an optimized set of components that do really good at a particular function. Presumably they want to design their GPU's so that developers of different applications (2D image editing, video editing, 3D rendering, VR, etc.) can have an optimzed set of architecture they can access for their particular needs.
For example, modifying a 2D image, where each pixel is a separate and unrelated calculation, likely requires less complex components and operations than, say, 3D scene rendering, where each light ray depends on a bunch of other stuff in the scene.
So faster components and different designs can make a big difference.
Also, not every task is designed for the thousands of parallel/simultaneous calculations that GPU's excel at. So conceivably, maybe the task you give the GPU just can't take advantage of all those parallel calculations, and at some point more cores doesn't improve things. For example if you have a 2D image with 10,000 pixels, but your GPU can handle 500,000 simultaneous calculations cuz it has a zillion cores, most of the GPU might be sitting idle, while a lower core count GPU might perform just as fast.
That being said, below is a summary of the render benchmark thread with rendering times (in minutes) for different models of GPU's. As you can see, two identical GPU's running in parallel doesn't mean you'll halve your render times. Also, take the numbers with a grain of salt, since different users' results can vary a lot for a number of reasons. Assume they are +/- 10 to 15 seconds.
I suspect if you add a core count to the spreadsheet the results won't directly match the expected render times based solely on core count. Though it would be great if someone could do some legwork and see if I'm right about that.
By the way, the middle column is the # of GPU's running, and the far right column is the render time in minutes.
By the way, I just did a quick check and the 980ti, for example, has 2816 cores, while the 1070 has only 1920, But the render times for the benchmark scene for both are identical at 3 minutes. Although I just finished a beer so I could be way off.
Okay, so I did the legwork and here's the updated chart with core counts included
Thanks a lot! Now I see the difference. And as I see doubling cores really matters, that was unexpected and kinda disappointing... starting to think of upgrade.
Not sure how you came to that conclusion ("doubling cores really matters"). From what I saw, an almost 50% increase in cores made no difference in the case of 980ti and 1070. Like I said before, it may make a relatively direct difference, but it's not the thing that tells you how two cards will perform. A 980ti vs a Titan X is only a small (12%) increase in cores, but render time dropped from 3 minutes to 2 minutes, which is a major 30% decrease.
And if you triple the cores from a 1060 to a 1080ti, you're only cutting render times approx. in half (a little better than that..), not in thirds.
It's a bit like comparing two vehicles' performance based soley on engine horsepower. Yeah, if the two vehicles are identical it might make a direct difference, but if it's comparing a Maserati with a 16 wheeler big rig, a tripling of horsepower may be irrelevant. A big rig isn't doing to do a quarter mile drag race in less than a few minutes no matter what HP the engine.
. It's about architecture.
Keep in mind that the vendors need to sell units, and they need their marketing departments to find something very simple to hype people into buying. So they come up with a single number that people can latch on to, mistakenly believing it's a direct indicator of performance.
"OMG, it has 5,000 cores !!! So everything is 5,000 times faster !!!!!!"
Yeah, you're right, but still significant change 1060 vs 1080, and when it comes about minutes there is not that much difference 10, 15 or even 20 minutes, but if render takes 8-11 hours it can be usefull.
I was expecting way smaller difference, something like 4-cores and 8-cores CPUs with 10-15% increase.
I've heard a new line of GPUs is coming 20XX, well, maybe prices will drop down so 1080 be like 300$
Within a generation your chart above shows almost exactly the expected increse in performance going from the 1060 to the 1070. From the 1060 to the 1080Ti does fall short, however. How much either of those owes to other aspects of the system is open to question - and of course there's no note of clock speed.
More cores within a generation generally translates to a significant, though not linear, improvement in performance from what we have seen. However, between geenrations core-counts are not the determining factor.
Yeah, I suppose with 8 cards to choose from you could sort the data however you want based on the desired result.
That's not what Richard was doing - he was trying to compare apples to apples (9 series) and oranges to oranges (10 series).
- Greg
And that was my point too...
IF your goal is to compare within generations, then you merely select the generation from the data and compare.
IF your goal is to compare ALL generations so you can make a general statement (as I believe the OP was intending), you'd select all the data.
You do indeed need to compare cards in the same series, and not across to make an informed decision. There are two main (and many other sub) reasons: First is that nVidia spends the bulk of their driver development time on their latest and current product offering, and for the GTX series, that's the 10xx boards.
Second is that the different series may use different designs and manufacturing processes that could impact speed. The CUDA cores themselves are actually fairly simple floating point processors, so (and this is nVidia's trade secret) those may not change much. But the gateway to the cores. and the GPU itself, will and does change.
nVidia has published charts showing the increase in GFLOP performance of their series of cards. If you're to believe them the increase is exponential, and surpasses Moore's Law. I'm sure there are some recent ones that compare the 9xx and the 10xx cards.
Yes, like I said. Drivers and architecture.