I’m running WIn10 on a core i7 870 2.93GHz with 16Gb and AMD HD6900 Series GPU 4GB DDR3.
I found no difference when choosing to render with GPU, seemed to use CPU regardless of settings as when I checked computer performance, in all tests CPU reported to run at 109% (The Device setting in Blender was also greyed out although I could change it)
CPU (16x16): 10:30
CPU (64x54): 8:40
CPU (128x128): 8:00
CPU (256x256): 6:36
CPU (512x512): 8:36
CPU (1024x1024) 11:15
GPU (128x128): 7:35
GPU (256x256): 6:30 - cpu?
GPU (512x512): 11:25
GPU (1024x1024): 40:00
Chances are your AMD card is not CUDA compatible, so no advantage.
http://www.amd.com/en-us/products/graphics/desktop/6000/6990#
Ha, you think your computer is old! Try the one I’m using! Actually, with CUDA, the old CPU does not seem to matter much (unless more complex scenes requiring RAM):
CPU:
Intel Core Duo 6600 @ 2.4Ghz x2
OS Lubuntu 16.04 LTS
GPU:
Dual GeForce GTX 750 Ti’s:
TL;DR:
GPU’s won handily over CPU (this was by design, special purchased for Blender)
Res-matched tile sizes beat out square tiles by a few seconds, which can add up.
In other words, for BMW27 benchmark scene:
Dual GPU’s won easily (10x’s as fast as CPU) with res-matched tile size:
240x135 *1:58.09
second best for GPU was using square tiles:
256x256 **2:00.53
best tile size for CPU was:
15x8 *23:26.63
second best for CPU was square tiles:
16x16 **23:30.97
=========Details============
Nominal 1920x1080
rendered at 50% resolution:
Square Tiles:
8x8 GPU 60+min(est) CPU 23:35.28
16x16 GPU 17:17.62 CPU **23:30.97
32x32 GPU 5:53.30 CPU 23:36.28
64x64 GPU 2:40.69 CPU 24:07.87
128x128 GPU 2:03.13
256x256 GPU **2:00.53 CPU 25:45.01
512x512 GPU 2:38.47 CPU 29.06.47
1024x1024 GPU 4:56.20 CPU 47+min (est)
Res Matched for rendering at 50% resolution
using following derived by
dividing full nominal full resolution by 256 to 2:
1920x1080 div
by CPU
/256 8x8 (23:35.28 above)
/128 15x8 *23:26.63
/64 30x17 23:32.84
/32 60x34 23:34.08
by GPU:
/16 120x68 2:17.96
/8 240x135 *1:58.09
/4 480x270 2:01.22
* = best in category (CPU vs GPU).
** = 2nd best in category (CPU vs. GPU) Res un-matched had slight disadvantage
Another thing to check is whether the Cycles Compute Device is actually enabled in User Preferences.
When I did my first tests I got ridiculously long render times as well e.g GPU 256/256 took around 6 mins and my device setting was also greyed out in Blender.
So I went to the user preferences again to double check and realised I actually had to click on my graphics card text! (possibly due to an alternative theme I have installed for Blender)
After that, my render time for GPU 256/256 is 1min 35s, which is a lot more reasonable.
PC specs: i5 3.10GHz, NVidia GeForce GTX 1060 6GB.
Before:
After:
