[dmeyer]

While we await the official new benchmark scene for GPU from NL, I ran some tests using the included library scenes. Note that some of the features used in these scenes are not supported on GPU, so be careful about making any conclusions about CPU vs GPU from this. However, its illuminating to compare the GPUs.

I cranked up the SL to the noted amount on these scenes, and disabled Multilight, image output, and Extra sampling to get a consistent read.

From the hardware I had lying around:

"Moet" Scene: SL 16
CPU 2x Intel Xeon E5-2699v4: 4:48
CPU 1x Intel Core i7 6950 @ 4.2 Ghz: 13:46
nVidia Quadro K6000 (Titan Kepler equivalent): 5:00
nVidia Geforce Titan X (Maxwell): 2:16
nVidia Geforce Titan X (Pascal): 1:33
nVidia Geforce GTX 1080 (Pascal): 2:07

"Beauty Care" Scene: SL 15
CPU 2x Intel Xeon E5-2699v4: 4:33
CPU 1x Intel Core i7 6950 @ 4.2 Ghz: 12:34
nVidia Quadro K6000 (Titan Kepler equivalent): 3:00
nVidia Geforce Titan X (Maxwell): 1:21
nVidia Geforce Titan X (Pascal): :54
nVidia Geforce GTX 1080 (Pascal): 1:15

"Guggenheim" Scene: SL 15
CPU 2x Intel Xeon E5-2699v4: 7:26
CPU 1x Intel Core i7 6950 @ 4.2 Ghz: 20:16
nVidia Quadro K6000 (Titan Kepler equivalent): 3:24
nVidia Geforce Titan X (Maxwell): 1:50
nVidia Geforce Titan X (Pascal): 1:11
nVidia Geforce GTX 1080 (Pascal): 1:46

While it varies by scene a bit. nVidia Maxwell is about 120% faster than Kepler, and Pascal is about 50% faster than that, so the newest Titan is over 3x faster than one introduced about 3.5 years ago. And the 1080 is roughly on par with the previous Titan Maxwell.

[Mihai]

Good info! Thanks for these tests. Keep in mind additive materials aren't supported yet in Maxwell GPU.

It would be great to see if the big difference between "Maxwell" and the newer "Pascal" architecture is also so big between the GTX 980Ti (Maxwell) and the GTX 1080 (Pascal). It might be worth investing in a 1080 rather than a 980Ti if the difference is almost 2x faster.

[dmeyer]

Good info! Thanks for these tests. Keep in mind additive materials aren't supported yet in Maxwell GPU.

It would be great to see if the big difference between "Maxwell" and the newer "Pascal" architecture is also so big between the GTX 980Ti (Maxwell) and the GTX 1080 (Pascal). It might be worth investing in a 1080 rather than a 980Ti if the difference is almost 2x faster.

I do have a 1080 I could test but don't have a 980 unless someone wants to send me one.

Also note that the above tests were using different CPUs, but I don't think that impact results at all considering they are not used much at all. They actually go into power save mode during the render.

[Neeper]

Wait - are you guys saying that a Titan Pascal is almost 14 times faster than a machine with dual E5-2699V4's?

That is a USD 1.200 graphics card vs a USD 12.000? machine? (CPU alone are more than USD 9.000) So to get to the same performance using CPU, you'd need a render farm of 14 of these machines. 14 x 12.000 = USD 168.000.

That is something like 67 times cheaper to get the same performance using a machine with Titan Pascal GPU. Way more than a 10x performance increase.

Am I missing something here?

Best,
Kim

[Mihai]

Dmeyer could add to this regarding his testing scene, but I think in the typical product/packaging shot scenes - the developers mentioned the speed gains could be 10x or more, using GPUs. So I think everyone is very eager to see the updates coming to Maxwell GPU so we can make more even comparisons with the CPU. Even for more complex scenarios like interior rendering, when Maxwell will be able to do Multi-GPU, buying a 600 card vs a new 3000 computer.....

[Mihai]

Looking at other tests with GPU engines it seems the GTX 1070 is not bad either compared to the 1080 which is few hundred $$$ more. The 1070 also has 8GB of RAM like the 1080 albeit a bit slower RAM and a bit less CUDA cores. But even so there isn't IMO difference to justify the much more expensive 1080. Might be worth it instead to have a 2x 1070 computer than a single 1080 one. This way you can also work and render on the computer at the same time if you want.

[dmeyer]

thread bump - rewrote initial post with more thorough test including 1080.

[Neeper]

Didn't see the original post was updated.

I ran the scenes, results on my 2011 MacBook Pro Core i7-2760QM (Quad 2.4GHz) upgraded with SSD.

"Moet" Scene: SL 16: 1 hour 15 min (GTX 1080 36 times faster)

"Beauty Care" Scene: SL 15 57 min (GTX 1080 46 times faster)

"Guggenheim" Scene: SL 15 1 hour 10 min (GTX 1080 42 times faster)

These times were estimates to get to the given sampling level after running for 5 min.

Time to get new hardware Why on earth has no one made a good looking 4K 27+" IPS PC display????????????????????????????????
RIP Apple.

[jfrancis]

Does anyone have a recommendation for which 1080? I've been looking at them but every time I make up my mind on one I read a bunch of one star reviews about it bursting into flames or otherwise rapidly dying.

[dmeyer]

Does anyone have a recommendation for which 1080? I've been looking at them but every time I make up my mind on one I read a bunch of one star reviews about it bursting into flames or otherwise rapidly dying.

For GeForce class cards I always go Asus or EVGA and have never had a problem.

[jfrancis]

1080 ti

http://wccftech.com/gtx-1080-ti-coming- ... vidia-aib/

[dmeyer]

Very curious to know if Maxwell's GPU mode relies upon FP32 or FP16 maths? The new Quadro GP100 features double rate FP16 and would be an absolute monster if FP16 is the focus.

[luis.hijarrubia]

We use all, but the FP32 the most. We use FP16 and FP64 in some strategical places, but not a lot.

[bitnaut]

Using the latest Mac OS X Maxwell Studio beta v4.0.0.18 on my trusty 2010 Mac Pro with same parameters as OP:

Moet: SL 16
OS X 10.11.6 El Capitan with 96GB RAM and 2 x 3.46 GHz 6-Core Intel Xeon: 19:45
Same system as above but using Titan X (Maxwell): 10:47

Beauty Care: SL 15
CPU: 11:30
GPU: 7:00

Guggenheim: SL 15
CPU: 14:36
GPU: 10:35

Now if only there a way to utilize both my Titan X's and not just one.