Pascal(1080Ti) vs Volta(GV100) vs Turing(2080Ti) vs Ampere(3080) on Lavina

[vladimir.nedev]

I am sharing some UNOFFICIAL performance tests that I have done here to test the performance of Lavina on the different nVidia GPU architectures.

Some important notes:
- the GPUs were tested on different machines, so this could have affected the results very slightly (Lavina is not really using the CPU)
- the performance difference depends very much on the scene
- the test scenes are the same we use to develop Lavina
- I could have made any kind of mistake while doing the tests, they haven't been independently verified, DO NOT base purchase decisions solely on these results
- FPS really means AA samples per pixel per second on a 1280x720 resolution

A summary of the test scenes, from left to right:

1. Evermotion scene with Ambient Occlusion instead of GI (lights disabled, reflections/refractions enabled, shading graphs for non-glossy rays only).

1. ao_I37_1_1sample - AO with 1 sample per camera ray
2. ao_I37_1_10samples - AO with 10 samples per camera ray

2. Evermotion scenes with default settings (2 GI bounces, lights enabled, light tree enabled, reflections, refractions, shading graphs for non-glossy rays only), except light cache is disabled.

1. E25_2 - exterior, Sun/Sky, no other lights
2. E28_10 - exterior, 45 rectangle lights, 25 sphere lights, 6 max spot lights
3. I30_5 - interior, 1 rectangle light (invisible), 2 simple portals, 9 mesh lights, 1 dome light (without texture)
4. I33_3 - interior, 1 rectangle light, 4 portals, 8 sphere lights, 1 mesh light, most area lights are invisible, 1 dome light (with texture), 6 V-Ray IES lights
5. I33_9 - interior, Sun, 15 rectangle lights, 2 sphere lights, 7 lights are invisible, 1 direct max light
6. I33_9_gi4 - same as I33_9 but with 4 GI bounces
7. I37_2 - interior, 7 rectangle lights (6 are invisible), 1 max direct light
8. I37_2_gi4 - same as I37_2 but with 4 GI bounces
9. I37_7 - interior, Sun, 18 rectangle lights, 3 sphere lights, 10 lights are invisible
10. I43_8 - interior, Sun/Sky, 8 simple portals, 2 sphere lights, 2 max spot lights
11. I44_9 - interior, 12 rectangle lights, 2 simple portals, 26 sphere lights, 21 lights are invisible, 60 max IES lights, 1 max direct light, 28 max spot lights

3. Evermotion scene with (clip) opacity enabled.

1. opacity_E13_9_no_lights - lights disabled
2. opacity_E13_9_lights - lights enabled

In summary:
- 3080 is anywhere between 6.2 and 13.5 times faster than 1080Ti
- 2080Ti is anywhere between 3.1 and 7.8 times faster than 1080Ti
- 3080 is anywhere between 1.7 and 2.1 times faster than 2080Ti



Greetings,
Vladimir Nedev

[Phillip Miller]

And here are the resulting averages of those 15 tests, normalized to the speed of the 1080 Ti:

[boris_furduj]

Hi, is better to take 2060 (30 rt core) or 3060 (28 rt core) for Vantage?
Thanks!

[vladimir.nedev]

I haven't tested the 2060 or the 3060 specifically, so I can't say for sure.
But most likely the 3060 will be way faster in raytracing based on our comparison of the 2080/3080 GPUs.
The 3060 also has 12 GBs of VRAM compared to just 6 GBs for the 2060 and more VRAM is also very important for a smooth experience.

Greetings,
Vladimir Nedev

[Phillip Miller]

Yes - what Vladimir said - a "30" class GPU is 2X faster than a similarly sized "20" class GPU.

[boris_furduj]

Thank you!!!

[Pixelo]

Hello !

What about the RTX 3090 ?
Anyone tried it with Vantage ?
Vladimir do you knwo perforfmance gain vs a 3080 ?

Thanks, have a nice day

[vladimir.nedev]

We haven't tested an RTX 3090. Considering it has the same type of memory as the 3080, you can probably extrapolate the performance of the 3080 based on the amount of CUDA cores both have.
So the 3090 has 20% more CUDA cores, so it's probably around 20% faster. It could be lower than that however. It also depends on the max boost clock, which depends on how well you are cooling the card.

The main advantage of the 3090 is that it has more than double the RAM of the 3080.
So if you have large scenes to work with, the 3090 is a good choice based solely on the amount of RAM.

Note that we also tested the A6000. It has 23% more CUDA cores than the 3080, but turned out to be only 9-14% faster than it.
However, the A6000 uses the lower bandwidth memory - GDDR6 (instead of the GDDR6X in the 3080 and 3090), so this could be the reason.

Greetings,
Vladimir Nedev