GPU speed sifferences

[MildMustard]

Well then there must be something specific with BVA's hardware setup on that machine maybe, as the 1080 should definitely be faster than the 980Ti and Titan X (Maxwell). We should test and compare another scene, are there any Modo Specific Benchmark scenes available (Interior)?

What you could try do is swap the 980Ti and 1080 in those 2 machines and run the test again too.

[savage309]

the 1080 should definitely be faster than the 980Ti and Titan X (Maxwell)

Keep in mind you are comparing the 1080 (180W) which is the 980 successor to the 980ti (250W) - the proper comparison is the Titan X Pascal vs Titan X Maxwell, 1080 vs 980, etc.
With that said, the 1080 is actually faster (or on par) on most of the scenes we have tested with Titan X Maxwell. We also check with the vendors to make sure the performance gains from the new architectures that we see are the expected ones.

Also, make sure you are using same drivers, SLI off, same settings and scene, that the reported time in Modo is correct (you can export to vrscene and test from standalone, I guess), etc.

Best,
Blago.

[BVA]

Keep in mind you are comparing the 1080 (180W) which is the 980 successor to the 980ti (250W) - the proper comparison is the Titan X Pascal vs Titan X Maxwell, 1080 vs 980, etc.
With that said, the 1080 is actually faster (or on par) on most of the scenes we have tested with Titan X Maxwell. We also check with the vendors to make sure the performance gains from the new architectures that we see are the expected ones.

Also, make sure you are using same drivers, SLI off, same settings and scene, that the reported time in Modo is correct (you can export to vrscene and test from standalone, I guess), etc.

Best,
Blago.

I tried everything I could think of..... I have put one card at a time into the same PC with latest drivers, same monitor etc. no SLI (as I only have one gpu in the system at a time). No matter what I do the 980Ti is just faster than the 1080.....always at least 20%. I would have expected the 1080 to be 30% faster.... unfortunately not.

Can you try increasing the GPU Ray Bundle Size to 1024 and GPU Rays per pixel to 16, then run the tests again ?
Vladimir Nedev

I tried these settings, the 980Ti went 10 sec faster, the 1080 17sec slower

It's possible that the result is scene dependent, so trying another scene with a different composition of lights and materials might be a good idea.
Vladimir Nedev

I will use one of our production scenes and do some tests this week and revert back to you.

[BVA]

I have run the same test scene with both the 1080 and 980Ti together. If I look at the screenshot attached, the blue is the 980Ti and the red the 1080 speeds.
If I make the right assumptions, the 1080 is faster in all areas but the cores are less...I assume that will be the reason for the slower renders?

[GidPDX]

Core # = linear speed up for CUDA?

That is not the case. The 1080 has many speed advantages. It's actually harder to say "that's a fast GPU" with confidence than it is a CPU IMHO.

If it were the case, you'd see the recent octane betas and other tests with the 1080 slower- but it's not. Many octane tests put the stock 1070 at just 6% less performance than the 980ti too. With the right 1070 you can easily overtake a stock 980ti when over clocking- and not by huge amounts. So the 1080 should be doing better in RT than what you're getting.

https://render.otoy.com/forum/viewtopic.php?f=9&t=56108&start=10


Do you have octane or the benchmark demo? You might want to download it and see how your card compares. That will help you narrow down if it's something with RT, or your setups/card. I look forward to seeing what other scenes end up doing too than just the headphones one.

I suspect things will change soon once GT uses the CUDA 8 kit to bundle but that's just my guess. Some say it will...some say it wont. If not then something is up for sure either with the scene or some other way RT is using the 1080 (or not).

[savage309]

Well, if you like to think that we are not doing something right with Pascal, this is up to you (we are not). CUDA 8 will not change the perf numbers, we tested that already.
Actually, Pascal is very similar rob Maxwell - you can read in technical details here - http://code-bg.com/blogo/?p=1494, but it is ~40% more power efficient. Which means that 180W Pascal should be roughly as fast as 250W Maxwell (180*1.4 ~ 240). Our test (and the ones in your link - Octane bench has 135 on TitanXm & around 140 on 1080s, which is 3% faster, which we had the first day 1080 was released - many, many months ago .....) are showing that 1080 is as fast as Titan X Maxwell (few percent faster), which is what everybody expects. Also because Maxwell and Pascal are similar, if a new compiler makes one of those faster, it should make the other one faster as well.
There will be always scenes and setups that are not in that range - that's why statistics looks at the average (or even better, the median).
The written above is not true for games - there are much more differences in the game hardware between the two architectures.

Is the scene available publicly somewhere ? It would be nice if it is so I can test it, add it to the nightly auto test (if it really is a corner case, we better monitor it closely), as well and report the problem with it to the nvidia team, so they could take a look as well.

Best,
Blago.

[BVA]

Is the scene available publicly somewhere ? It would be nice if it is so I can test it, add it to the nightly auto test (if it really is a corner case, we better monitor it closely), as well and report the problem with it to the nvidia team, so they could take a look as well.

Best,
Blago.

Blago, what is your email address, I will send you the scene.
It is one of Richard Yots tutorial scenes we used to compare internally. (with very little changes, so still modo materials etc)
I did run some other benchmarks too. Indigo and Arion is faster on the 1080
Octane is also faster on the 980Ti vs the 1080.

I think I made my mind, if I do get more gpu's I will get the 1080's. they do render a little slower than the 980Ti's but use less power, probably generates less heat too. Thanks for the help/ feedback on this

[rvectors]

[ATTACH=CONFIG]33144[/ATTACH][ATTACH=CONFIG]33145[/ATTACH]

PLEASE can someone explain to me what can cause the speed difference?
I have used 2 identical machines, (different motherboards but otherwise almost identical. One have a gtx 980Ti and the other have a 1080.

The 1080 rendered 8m2.3s and the 980Ti completed in 3m53s

That is less than half the time on an older card? I have also picked up that denoising is a lot slower on the 1080 than the 980Ti.

Both are on 372.70 nvidia driver, both on windows 10.

That's interesting, if I recall correctly a 1070x1 (372.70) with the same scene in default, topped out at around 6 mins. The CPU RT just over 12 mins.

Prod (CPUx1) on the other hand, less than 3mins!


Not withstanding the better speeds mentioned in Indigo etc, I did read somewhere that the 1080 may have issues with the 'X' variant memory but the latest drivers were meant to address this. Then I 'thought' I read somewhere else that Nvidia may actually be physically limiting the memory via the driver, for stability reasons. All of this is read on the internet, so pinch of salt...


Slightly off topic
As I understand (in my simple evaluation stage/inexperience), the RT is locked to progressive, so this would explain the prod vs RT performance? BVA's pictures also remind me of something when I was testing a while back, the GPU scene is missing specular highlights compared to cpu. I'm sure those full time users don't find this odd but the RT compatibility list didn't point to anything obvious. It seems like the GPU RT is not using the light cache, even though it calculates it?

[GidPDX]

Octane is also faster on the 980Ti vs the 1080.

That should change as soon as they've updated octane, at least that's what they keep saying.

Some beta testers have had good results with 1080s and 1070s. Not to sound like a skipping record, but if in some tests a 1070 can = a 980ti in Octane then I'd expect the 1080 to on average be as good if not better than the 980ti. Obviously this is all before anybody overclocks things and on certain scenes etc.

Sorry if I spread mis-information about the possible performance gains with CUDA 8 vs. 7.5 with RT. I just (wrongly) assumed that could be the case given how vastly slower the 1080 was on that particular scene and the claims made by Nvidia and some other render engine developers.

Hopefully CG can figure out what about that headphones scene was causing the difference and fix it.

BVA- why is the 1080 render higher res than the 980ti one?

[savage309]

That should change as soon as they've updated octane, at least that's what they keep saying.

I was left with the impression from your previous posts that Octane is much faster with 1070 compared to 980ti .. apparently this is not the case but it will be in the future, my bad..
If they say so, it has to be the case, but if somebody can explain how a particular optimization could affect GP102 (to GP107), but will not have the same effect on Maxwell (beside doing this on purpose), considering this http://code-bg.com/blogo/?p=1494, I would be really thankful ...

Best,
Blago.

[GidPDX]

No worries!

Currently in octane the 1070 is often very close with a 980ti in the benchmarks on the beta builds. This is with a 1070 FE at stock speeds.

https://render.otoy.com/forum/downlo...5789&mode=view

What I did say is that an over clocked 1070 can overtake a 980ti (not over clocked)- but that was conjecture based on tests outside of octane which show it getting results closer to a 1080. So who really knows until more testing on more scenes is in.

My point wasn't to say so much that the 1070 is faster/as fast as a 980ti, but that if in octane tests it is close to are equal to a 980ti, then we should expect for cuda rendering on most scenes, that the 1080 should be faster on average, if at least not equal.

Regardless, I'm REALLY hoping that BVA's issues he's having are something particular with that one scene and not Vray for MODO in general. I expect is the just that one scene, but until we see more tests with MODO scenes running in RT we can't say. I think everyone agrees that a more than 40% difference between the 1080's and the 980tis on his tests is strange.

I'll probably buy a few 1070's or 1080's as I'm not interested in 980tis due to the heat/power/RAM limitations, but I'd hate be using them only to denoise in Vray, and only take advantage of them for rendering in other engines.

[savage309]

take advantage of them for rendering in other engines.

I can't explain what I said better than I already did here.
If you happen to take advantage of them only in other engines, let me know.

Best,
Blago.