CPU: Intel I7-7700k @ 4.2GHz
GPU: 2x MSI GeForce GTX 1080 Aero 8G OC
RAM: 32GB (4x8GB) Corsair Vengeance LPX DDR4 3000MHz C15
SSD: Samsung EVO 960 M.2 1TB
MOBO: ASUS TUF Z270 Mark 1
PSU: SeaSonic Snow Silent 1050W
OS: Windows 10 64-bit
As you can imagine, GPU rendering is a no brainer here. Still, testing was done across both methods. I will save you a LOT of reading just get to the good stuff.
GPU rendering 1920x1080 at 50%
We are only testing for tiles sizes here.
I found the best 2 tile sizes to be 480X/540Y and 240X/180Y. Without getting deep into the math on it, these numbers just work. I ran a series of 7 rapid renders on each setting to check GPU temps and render speeds.
(Idle temps are 40/38)
480x540
[Render Time] @ [GPU1/GPU2 Temps] (in Celsius)
0:33.53 @ 53/47
0:35.17 @ 58/55
0:34.70 @ 60/60
0:39.79 @ 61/61
0:40.30 @ 63/59
0:34.81 @ 63/63
0:34.88 @ 64/64
240x180
0:34.61 @ 55/54
0:35.47 @ 58/57
0:34.93 @ 59/58
0:34.87 @ 61/61
0:34.87 @ 62/61
0:35.21 @ 63/63
0:35.18 @ 64/61
What’s interesting is that while 480x540 rendered faster than 240x180 in a single instance, the variance of render times was quite extreme. 4.77 seconds may not seem like a lot, but with this being a single frame, all those seconds matter if this is an animation. 240x180 only varied by 0.86. This also means the average render time favors the smaller tiles with 480x540 @ 36.17 seconds and 240x180 @ 35.02. Fun Fact: The Auto Tile Size Add-on selected 240x180.
I was convinced that splitting the render into 2 and dedicating each card to 1/2 would have been the best option considering which cards are being used. These results definitely surprised me. I did the same test MANY more times and found the results mostly reinforced the results above. The stability of the smaller tile size makes it a better option overall. That being said, if you wanted to roll the dice just for speed, then you could get better results from the larger size. Emphasis on COULD.
Fastest render times from ALL tests
480x540 @ 0:31.63
240x180 @ 0:34.22
You could also get far worse results…
Slowest render times from ALL tests
480X540 @ 0:46.56
240x180 @ 0:36.18
So what about using the CPU? Not a chance.
To be fair, the CPU did beat out the GPU in certain situations, but overall, it can’t compete.
Just a taste:
CPU vs GPU fastest renders
240x180 (selected by Auto Tile Size for the GPUs)
CPU @ 3:03.11
GPU @ 0:34.22
32x32 (selected by Auto Tile Size for the CPU)
CPU @ 2:29.21
GPU @ 4:00.98
Side note: I was glad to see that Blender Cycles utilizes up to 100% of both GPUs simultaneously. No scaling issues that you would get from SLI. This makes me wish I had shopped around for an EATX mobo so I could slot in another GPU for Cycles rendering.
Hope this helps someone. Take care and have fun!