Complex Interior Visualisation: GPU perform much worse?

[cadman5]

Hello all:

I did a (1500x1000px) test of an interior office lobby scene (16 unique lights+ many light instances, 28 million polys). The system itself has 128GB of Ram. Using Vray Advance 3.60.01

Render time with a dual xeon e5 2697 v4 was 5 minutes!!

Render time with Vray rt gpu and a single gtx 1080ti was 1 hour 50 min!!!!!

Both noise levels were set to 0.005

Also, the Vray advance image was noticeable cleaner. A second gtx 1080 ti card will only improve my result in 1.8 times according to some websites, and for every 4 cards we will effectively "lose" the equivalent performance of one card...

Am I right considering that GPU rendering is not comparable with 2 good xeons (or even the 1950x threadripper) for complex interior scenes?


Thanks for your time reading this and appreciate your input.

Cheers

[matanov]

Hi,

This is not normal. Please send me the scene to check it here. There could be something in the scene that slows the gpu down.

[cadman5]

Thanks Peter. Wetransfering the file to you right now. Cheers!

[matanov]

The sent scene was set with V-Ray GPU as render so I don't know what were your CPU settings so I rendered with the default with only the noise threshold set to 0.005

Here are the results:

i7 5820 BF LC - 1h 19m
i7 5820 IM LC - 30m

1xTITAN X (Pascal) BF LC - 27m 22s
1xGP100 BF LC - 14m 8s
2xGP100 BF LC - 7m 40s

All render with V-Ray Next beta 3.

[glorybound]

WOW, 7 minutes compared to over an hour? Now, the 7 minutes will cost you $15,000, but impressive.

[cadman5]

Hi Peter:

Thanks for your tests.
I have attached here the vraycfg preset file I used if you would like to test the scene with my CPU settings. My render time with a dual Xeon 2697 v4 (vray 3.60.04) was just about 5 min.
I guess there should be something wrong with my GTX 1080Ti since it took 1 hour 50 min here and your Titan X finished in 27m 22s. (over 4 times quicker)

That aside I noticed from your test that even 4 Titans on Vray Next GPU wouldn't be faster than the 2 mentioned Xeons on Vray advance 3.60.04 using my settings (IM+LC) for complex interiors.
And considering that hardware costs will be similar + more than double the energy bill using cards It just means It won't make sense economically speaking either. Thanks.

[super gnu]

despite being a longtime supporter of the irradiance map ( particularly for flytrhoughs) you cant really compare to a BF gpu render. the quality of the BF render will be much better in most cases (increasingly so in difficult ones) and requires 0 tweaking or setup time. grain is far preferable to blotches, and can be cleaned wonderfully with the denoiser. if you got to the bottom of why your 1080 is on the go-slow, it should (you can even add your cpus to the gpu rendering job) give a decent benefit. ive swapped over to GPU almost entirely now (but still curse the measly ram limit for larger jobs) as it often renders many times faster than my 6 core 4.6 ghz cpu. i only have one titan X.

by the way, did you try precalcing the imap and rendering with it in the gpu engine? (unless they removed that feature again)

[cadman5]

Hi super gnu:

I get what you are saying about quality. But I am more interested in raw speed with 0 noise for many daily tests, not for final rendering. As it turns out, I can achieve the same if not better result with the two Xeons (72 cores) faster than 4 Titans X if we are going to extrapolate the results of Peter's tests.
BTW, I never get what some people say about having to do a lot of tweaking with Vray Advance. I never do any tweaking and most of the time I don't even use the denoiser.
You can check the settings I use from the vraycfg file I attached before. Nothing special at all, they just suits my needs fine.
I am aware of the appeal of unbiased "pure" render , the "superior" GI detail of brute force on GPU and all that.
But I have to say the designers I work for don't appreciate any difference anyways. But man, they do hate even the slightest noise on the tests..

[glorybound]

That link you posted doesn't work for me.

[cadman5]

Its a rar attachement containing a vray.cfg preset file.

[glorybound]

I get this
This site can?t be reached

The webpage at https://forums.chaosgroup.com/filedata/fetch?id=994842 might be temporarily down or it may have moved permanently to a new web address.

ERR_SPDY_PROTOCOL_ERROR

[cadman5]

I just downloaded it to check and it works. Try a different browser or disable anti-banner feature.

[matanov]

I can see from the CFG that your CPU settings are far from the defaults that I tested with. A proper comparison between the engines, as super gnu said, will be fair only when using BF + LC.

I've also ran the scene with 3.6 with the max samples set to 5000 (=the new default) and the result was 42 minutes, that means V-Ray Next beta 3 is around 60% faster in that scene (our GPU team rocks ).

Bottom line is - it's a good decision we changed the default max samples value, because 10000 tend to oversample the image (your case).

[cadman5]

Peter,

The reason Vray became the most popular renderer was the speed. And that speed was possible because Vlado implemented algorithms that gave brilliant results in less time (shortcuts as many used to call them). The were not perfect, but almost perfect and good enough for most people.

I am sorry to disagree but there is not a "proper comparison" here. If I can achieve 95% of the Brute force GPU quality at half the time or less without noise, that is more than good enough for me.

Vray GPU is twice as fast as Vray RT but as I see it, it is just one of the many choices out there, not particularly magic as Vray Advance has been for many years.

As I said, thank you for your tests. If I were to chose a new a hardware now, for the same price there is no doubt in my mind I would chose a top notch CPU over GPUs for complex interiors.

[ramon_perez]

There's no way a sinlge GTX 1080ti can outperform a high end dual xeon setup.

4x 1080tis in the other hand can compete with the xeons any day for half the money. Wait for the next generation 20xx series and see what happens.