GTX 780ti only marginally faster than CPU?

I recently installed an Asus OC 780ti into my Dell T5610 running dual Xeon E5-2630v2 at 2.6ghz (12 core total) and was surprised to find that while it rendered faster than the K2000 that came stocked (expected) it was only marginally faster than the CPU (1:05:30 v. around 55s) and certainly no faster than other benchmarks using 500-series geforce cards.

I think there was a power issue as the PSU shut down on a stress test, and will revisit with a dedicated GPU PSU I have installed and waiting on a SATA to IDE adapter to get it going… But am I expecting too much? Should I just sell the 780 and maybe get a K4000 instead?

Hi,

I think the K4000 might be better for viewport display, but not for rendering.
On what file did you tested your rendering ? with which parameters ?
I assume you already know that tile setting is critical when you want to optimize render time and that CPU perform well with low tile size (8x8 or 16x16) but GPU perform better with large tiles (256x256 or near).

Anyway it is a powerfull CPU configuration you got there, maybe you need some challenging render to notice the difference.

Actually, I don’t know a lot about optimization, and the whole GPU thing is pretty new to me.

The file I used was the BMW benchmark which I got from this site via google - I am sure you are familiar with it. The benchmark was 1:1 on both CPU and GPU with identical settings. I will try optimizing tiles on a more complex scene and see what kind results I get, of course, the card may have been power starved, also, since I am not sure exactly how Dell is managing distribution. The card is outside published specification for the T5610.

The adapter for the dedicated PSU should be here today in a couple hours, I’ll do a more indicative test then.

Once the tile setting was set to 256x256, I saw a 50% increase in performance. Render time went from about 55sec to about 30sec, and a less dramatic improvement to CPU performance when tiles were set to 8x8.

So yeah, that appears to explain the issue. At 128x64 tiles the performance about matched, but once optimized the GPU more than halved render time.

The only problem is that this particular card draws too much power, it really is too big for the case and the dedicated PSU I have is too noisy and too hot. So I will be needing to trade it out for a 780ti reference unit.

GPUs prefer larger tile sizes… you may see a speed increase going up to 512x512 (provided the render size is large enough)…

I have found for a 1280x720 render, 640x360 is usually the fastest tile size for gtx590s… May differe for yours.

Definitley need to play with the tile settings.

I will admit that I have never gotten anything from rendering at 512x512 compared with 256x256 on my GTX680, but it might be an architectural difference. in some testing the GTX 570 was on par with my GTX680.

would be interesting to the Tile setting automatically setup depending rendering device. not each setting for each GPU, but that when rendering on CPU, auto set to 8x8 or 16x16, and on GPU auto set to 256x256. Most novice users will automatically see speed increase without dabling with settings.

Here you are:

http://wiki.blender.org/index.php/Extensions:2.6/Py/Scripts/Render/Auto_Tile_Size

Cheers, mib