Nvidia Maxwell GPU Benchmarks

I FINALLY FOUND SOME!!!


this is the original site:

it looks like the 970 is on par of the 780Ti on a much cheaper price!
for any adopter that are looking into upgrading, this is your chance!
since GTX 980 is only for $499

great card only downside is 4gb vram only

i believe that they are going to release 8GB variants later on, if not can you not just SLI it?

Memory isn’t combined when you SLI cards.
I have a 970, if you want i can post some benchmarks for you?

Since there are no official builds with sm_52 capability around and I highly doubt that Sweclockers has the time and ability to build Blender just for this test, I would not take these results for granted to say the least.

yes please, and please tell us if the benchmarks are about right

According to BeerBaron on IRC (who, sadly, is apparently unwilling to register on BA) and the CUDA C Programming Guide, you should be able to rename the sm_50 cubin to sm_52 and it should work.

Of course, disclaimer, if your computer bursts into flames it’s not my fault, blah blah. :smiley:

That said, it’s really not that difficult to compile Blender from source (even in Windows).

Alright, since someone went through the trouble of registering an account for me, I shall grant you the precise instructions that Fweeb was somehow unwilling to give you:

  1. Go to your Blender installation directory, for example ‘C:\Program Files\Blender Foundation\Blender’
  2. Navigate to ‘2.71\scripts\addons\cycles\lib’
  3. Copy kernel_sm_50.cubin and rename it kernel_sm_52.cubin

This should work just fine, since according to the CUDA specs, there is binary compatibility between minor versions.

Unfortunately I’ve never compiled software before. If someone would be willing to compile it for me that would be greatly appreciated.

Renaming the files as described above works (for me). At the very least I can render things with a 980- Compiling blender from source is only “not difficult” if you already know how the tools work. There are several holes in the documentation that an average user can’t fill on his own.

You don’t need to compile anything, just rename the cubin, like I explained above.

double post, sorry

oops, accidentally missed your post. Got cycles working now.

A quick test using mib2berlin benchmark file gets a render time of 3 mins 19 seconds with stock MSI 970 settings. I haven’t had a chance to overclock the card yet but the MSI cards are supposed to be really good overclockers.

Considering that 2x GTX670 achieve 02:48:27 this seems like a good time.

So what is the purpose of changing the file name?
Is it reason specific to the 970/980s or is it Maxwell in general?

Blender already ships a binary for earlier Maxwell GPUs (GTX 750 and 750Ti with compute capability 5.0) named kernel_sm_50.cubin. Blender queries the GPU compute capability (5.2 in the GTX 9xx) and looks for a file that matches it by name. Since Blender doesn’t look for anything but an exact match, you have to copy and rename that file to kernel_sm_52.cubin, or else it won’t find it.

Cool, thanks.

Very true. The building blender wiki page that everything links to does fine till about 70% of the way through the process then it just trails off. I spent about a week trying to get blender to compile, finally managed to get it to almost work, but I couldn’t compile CUDA and 64 bit without Visual studio pro. Dealbreaker :frowning:

Here is a Windows x64 build with a real sm_52 kernel: http://blender.dingto.org/blender_win64_geforce_maxwell.zip (Only contains sm_52 kernels, no other archs).

No idea if this dedicated one is faster than a renamed sm_50 one, happy testing. :slight_smile:

Edit: CPU rendering in that build has a bug, ignore it, already fixed in master.

Unfortunately I’m missing VCOMP120.DLL, so i’m unable to open that build. How do i replace the missing library because everything I have found points to it being a library from Visual Studio?