980 Ti vs Titan X vs 1080 vs 1080 TI Benchmarks

[Donfarese]

I have had some questions about just how good the new 1080Ti is and what gpu's to buy next. I Wanted to upgrade my 2 980 Ti's to something faster and with more vram. The benchmarks I ran with the cards I purchased was surprising. I ran benchmarks with the same scene in Vray RT and Redshift, all cards where tested in the same computer. Check the graphic for benchmarks. I also attached the scene for you guys to test yourself, all you have to do is change GPU selection and render. Both Vray RT and Redshift scenes are in the zip.

I noticed that the render times in Vray are quite weird, like they don't utilize the Pascal video cards correctly. Especially the 1080 which renders really slow for some reason. I also noticed now that I'm done with all the benchmarking that there is a thread by Blago saying not to use the 380+ drivers that of course I used. So not sure if that's the complete issue.




Sorry I don't know why I cannot get this graphics to show up in full resolution, and how to get rid of that Photo Album thing at the bottom.

[savage309]

1080 should be roughly as fast as Titan X Maxwell (plus or minus a few percent). 1080ti should be 30%-40% faster than 1080. In our tests, for most of the scenes that's the case.
Drivers 380 (probably after 378 ) can cause some troubles (crashes), not sure if they affect the performance though.

Make sure to use V-Ray GPU as production render. The production settings are optimized to utilize the GPU well, while the Active Shade settings are optimized for faster feedback. There should be no issues with utilizing the GPUs (if there is, something is wrong). Btw, it looks like that either the V-Ray render has more GI light or that the V-Ray scene is brighter in general.

Best,
Blago.

[glorybound]

What about MAX? It seems MAX only performs well with the quadro video cards. The quadro video cards seem to be 2X the cost for 1/2 the speed. I think I made some mistakes buying cards that work good with V-Ray, but poor with MAX.

[Donfarese]

1080 should be roughly as fast as Titan X Maxwell (plus or minus a few percent). 1080ti should be 30%-40% faster than 1080. In our tests, for most of the scenes that's the case.
Drivers 380 (probably after 378 ) can cause some troubles (crashes), not sure if they affect the performance though.

Make sure to use V-Ray GPU as production render. The production settings are optimized to utilize the GPU well, while the Active Shade settings are optimized for faster feedback. There should be no issues with utilizing the GPUs (if there is, something is wrong). Btw, it looks like that either the V-Ray render has more GI light or that the V-Ray scene is brighter in general.

Best,
Blago.

Hi Blago, this wasn't a render speed test against Redshift, I used both to make sure I get the same render times percentages against the cards to get a more accurate idea of the differences against the cards. I used each card to render in production render of course, there is no way to get accurate results in active shade. Check and try the Scene I uploaded. the best people are seeing over at Redshift is an increase of around 20% from a 1070 to a 1080 TI. I know it sounds crazy but it's true from there tests and look at the render times I get also even on vray. I see you are getting 30-40% better timeswith the 1080 Ti, but I'm surprised it's only 20-28% faster than 980 TI which means the 1080 isn't that great of a compute card. Each gpu used 100% usage and didn't throttle from what I can see. Not sure what the problems are from the vray render times especially with the 1080 being so slow. are you sure it's not something in vray not utilizing the Pascal right for some reason? again try the scene so we can compare.

[Donfarese]

What about MAX? It seems MAX only performs well with the quadro video cards. The quadro video cards seem to be 2X the cost for 1/2 the speed. I think I made some mistakes buying cards that work good with V-Ray, but poor with MAX.

No you are talking about viewport performance, Quadros are better at that, but for a much larger price that you might not see any difference in real world results. For Gpu rendering Geforce cards are better and faster, unless you spend 8-9k on a P100 or whatever it is. But then again you can just buy two 1080's for 1k and beat that.

[Donfarese]

Blago, Have you had a chance to test the scenes with your video cards to see the difference and why certain cards are getting horrible times?

[glorybound]

No you are talking about viewport performance, Quadros are better at that, but for a much larger price that you might not see any difference in real world results. For Gpu rendering Geforce cards are better and faster, unless you spend 8-9k on a P100 or whatever it is. But then again you can just buy two 1080's for 1k and beat that.

I know, but if you are spending 3/4 of your time building the scene in MAX that is important.

[savage309]

The Pascal architecture for computing is completely the same as the Maxwell architecture. Pascal GPUs however are about 40% faster for the same power consumption because of the more efficient process/transistors (14nm finfet vs 28nm planar). The only exception is GP100, which is quite different to all other Pascal architectures.
Therefore, it is almost impossible for us to do anything wrong for 1080, but right for Titan X Maxwell. The 1080 TDP is 180W. Titan X Maxwell is 250W. 180W * 1.4 = 250W, meaning it is expected 1080 to render as fast as Titan X Maxwell. Which makes 1080 a great GPU - it is actually 980 successor, it is much cheaper than Titan X Maxwell was, uses less power and is just as fast . Does this explanation makes sense / answers your question ?

Best,
Blago.

[Donfarese]

The Pascal architecture for computing is completely the same as the Maxwell architecture. Pascal GPUs however are about 40% faster for the same power consumption because of the more efficient process/transistors (14nm finfet vs 28nm planar). The only exception is GP100, which is quite different to all other Pascal architectures.
Therefore, it is almost impossible for us to do anything wrong for 1080, but right for Titan X Maxwell. The 1080 TDP is 180W. Titan X Maxwell is 250W. 180W * 1.4 = 250W, meaning it is expected 1080 to render as fast as Titan X Maxwell. Which makes 1080 a great GPU - it is actually 980 successor, it is much cheaper than Titan X Maxwell was, uses less power and is just as fast . Does this explanation makes sense / answers your question ?

Best,
Blago.

Like always thank you for your explanation Blago. It makes sense, just was thrown off because quite a few people over at Redshift I see were also benching there 1070's against TItan X Maxwells and getting the same times +/- 5%. So to find out that the 1080 is up to say 19% slower than Titan X Maxwell was quite a surprise, even though it was kinda expected. I mean 1920 CUDA Cores vs 3072.... But anyway it was great to get info from you also. My 980 Ti's still seem to be the best bang for the buck, but they are on the lower side of vram now a days. How many Polys can say a 980ti with 6gb Vram handle roughly?

[savage309]

The performance of the GPU depends on the cores speed as well. Also, on some scenes 1080 will be slower, on other faster than Titan X Maxwell. On average, we found them to be roughly equal (although the test was made some time ago).
If I remember correctly, we are using about 250MB for each million of polygons, more if you use motion blur or need extra channels (for example, some types of anisotropy require those).

Best,
Blago.

[Donfarese]

The performance of the GPU depends on the cores speed as well. Also, on some scenes 1080 will be slower, on other faster than Titan X Maxwell. On average, we found them to be roughly equal (although the test was made some time ago).
If I remember correctly, we are using about 250MB for each million of polygons, more if you use motion blur or need extra channels (for example, some types of anisotropy require those).

Best,
Blago.

We where going to be using a scene that we created in Redshift originally but are converting over to Vray cause all the assets we need to use are in Vray materials. Redshift can handle 1 million polys to every 60mb, so we are a little worried if it will fit everything we need in Vray RT. Also I did a test yesterday in Vray RT with a sphere with 70k polys and duplicated it with just copy, not instance to around 64million polys and that rendered, so I'm a little confused to the 250mb to 1million polys. Because that would mean I can fit 24 milliom polys at best on my 6gb cards, not the 64 million. Unless I was going out of core and didn't notice.

[savage309]

I think we have made some calculations long time ago. But it was really long time ago, so I might have as well just forgot what the actuall number is I think we should use somewhere around the bare minimum for geometry, so you should be fine.

Best,
Blago.