1070 vs 580 and 2.77 vs 2.78

Just spent a morning benchmarking my new 1070 over my old pair of 580s and thought I’d share the results if anyone interested:

I used the Blenchmark single car render scene and for each test varied the number of tiles until I got the best performance (normally ended up between 256x256 and 316x316).

I used the latest build of 2.78.

Here are the stats - only showing the best times obtained from the sweet spot number of tiles:

1070 on it’s own: 32 seconds

1070 + 1x580: 22 seconds

1070 + 2x580: 18 seconds

1x580 on it’s own: 56 seconds

2x580: 31 seconds

And then out of interest I went back to version 2.77a to re-benchmark the 580s and found they where slower with this version:

1x580: 70 seconds
2x580: 42 seconds

The really interesting outcome seems to be that version 2.78 significantly improves the speed of the 580s - to the point that 2 of them give the same performance a brand spanking new 1070 (although clearly without the same memory).

Forgot to mention running on Windows 10

That is interesting, though with only 1.5gb VRAM it’s a bit misleading, not to mention the cost in terms of power consumption. You are greatly limiting your 1070 when you attach the 580s onto it, though if your pipeline is trim it will work just fine for you.

I have noticed a slight decrease in CPU performance and a moderate increase in GPU performance with 2.78 over 2.77.

I should have made that clearer - don’t want to misslead anyone (thought people would know it had only 1.5gb (or 3gb if you have those versions)).

I’m planning on keeping one in my system to use in conjunction with the 1070 when I’m playing around learning with small scenes. When I need more memory I can just switch to just using the 1070 in the preferences. (I believe there would not be complication with the 580 in the computer as well).

Wow Thank you!! I have a gtx 590 and now I know what I MUST buy to keep good performance.
Unfortunately for me i only have 1.5gb per card and some of it is used for display.
Does someone knows how a 1070 perform versus a rx480?

I think 1070 is your best card at the moment. I don’t know how it compares to the 480, but the Pascal cards have decent OpenCL performance along side their proprietary CUDA, which blender supports well and is used by a number of other applications as well.

I am not a fan of CUDA, but it is kind of a necessary evil if you’re into rendering on the GPU.

BTW - in the meantime you can sometimes use some render layers if you need to render a large scene with lots of objects, provided that no one layer requires more memory than is available.

Thanks for the advice Shawn. I have to do archviz with blender and sometimes a simple tree or grass takes the whole memory avaible ( about 800mb). So even with render layers, it’s hard for me to manage every layers.(I am not good enough).
I was thinking about the price per second of a rx480.

How are you managing to get 32 secs for 1 1070? I just got two and I’m getting 40 secs to render the BMW blenchmark scene using the default settings.

I’d be curious to see the results from Linux, compared to Win10. Do you have access to a Linux liveOS or similar?

Why still benchmarking the old BMW scene while there are a new scene and a thread for this?

256x256 should not be the best resolution for a GTX 1070 alone for symmetric reasons.

Here is my result on a Gigabyte G1 GTX 1070.
Blender: Daily build https://builder.blender.org/download/
OS: Ubuntu Linux 16.04
Tile size: 480x540
Render time: 00:20.25

Please notice that my Intel Haswell CPU turbo is only 2.6 GHz.

Nice I’ll play around with the tile size settings - you’re getting about half of my time and I have two 1070s :eek:

Hi, keep in mind there are several “BMW” .blends around, this seams to be the very first one.
There is BMW27.blend with 400 samples and a BMW27.blend with 1225 samples and one with 1600 samples.

Cheers, mib