Recommendation - Time to upgrade, but what? powerful GPU, multiple GPUs, more CPU?

Support Technical Support
1

Hi all,

I’m a Blender hobbyist (for now) and I’m considering upgrading my computer. So, I’m thinking about upgrading my graphics card, but I am not sure which makes more sense in terms of a Blender workstation… I’m trying to work out the best bang for the buck and would appreciate any advice. If or when I do animation stuff, I’ll probably use a farm… so my question is really about working in Blender on a daily basis.

Background: I am using a Blender 2.78 under Linux. My current machine is a dual quad-core Xeon 3.2Ghz (8 real cores) with 24GB ram… My main GPU is a Nvidia GTX 680 2GB card. (my mainboard is only PCIe v2 - not sure if that matters) In my non-scientific tests, switching between my 8 CPU cores and my GPU seems to yield similar render speeds (GPU is a couple seconds faster per frame on average)

So… I’m trying to decide which of these makes the most sense:

  • Get two GTX 1050 ti 4GB cards. Replace my GTX 680. Total cost: ~$300 usd
  • Get one GTX 1060 6GB card, use 680 for display. Total cost: ~$270 usd
  • Splurge - Get one GTX 1070 8gb card, use 680 for display. Total cost: ~ $400 usd
  • Off-the-wall option - Save my money, get a dual 10-core Xeon (40 total CPU threads) and do it all on CPU. Total cost: ~$1k usd

Does the difference between 4GB on a GPU and 8GB on a GPU really that big? I have read about compositing larger scenes (I’ve not made any that big yet) but it seems like compositing + multiple fast-ish cards would be better than one really fast card with more RAM. Is it likely, as a hobbyist, that I’ll even run against the limits of a 4GB or 6GB card? Would I be better off just sidestepping the whole memory limitation issue and invest in more CPU horsepower?

Any thoughts welcome!

Jay

2

In your situation, I’d go with the Option 3 or 4. The GTX 1060 6GB currently has the best price/performance of NVIDIA cards. I wouldn’t invest into CPUs until Intel has responded to AMDs new CPU lineup.

Does the difference between 4GB on a GPU and 8GB on a GPU really that big? I have read about compositing larger scenes (I’ve not made any that big yet) but it seems like compositing + multiple fast-ish cards would be better than one really fast card with more RAM.

Compositing scenes like that is extra work and kind of defeats the purpose of having a physically based renderer, where all objects in the scene correctly interact with one another. I’d want to avoid that, unless it’s necessary.

Is it likely, as a hobbyist, that I’ll even run against the limits of a 4GB or 6GB card? Would I be better off just sidestepping the whole memory limitation issue and invest in more CPU horsepower?

Even as a hobbyist you can easily create scenes that exceed your memory budget. It really depends on what you want to do. If you have more time than money, you can spend the time optimizing your scene. If you have more money than time, you can push the baseline to where you don’t need to worry about it as much. Either way, in 2017, you should go at least with 6GB of VRAM.

3

I have a GTX770 GPU for now, and i’m also thinking in change it for a GTX 1060 or GTX 1070… but in blenchmark there is no great difference between the 770 (82 seconds) and the 1070 (71 seconds)

The same for your GTX 680 (88 seconds)

http://blenchmark.com/gpu-benchmarks

But I think that list is not fine, because I render in 132 seconds the 2 BMW scene … not in only 82 seconds…

How about you ? Anyone have a GTX 1060 or 1070 to give us the real time test ? Your GTX680 do it in how much time ?

Thanks to everybody ! :slight_smile:

4

Forget about blenchmark, the tile sizes tested are way too low for GPUs, especially the newer ones. Your best bet is probably digging through the BMW benchmark thread or the comments section of the BI tests.

5

Thank you ! With those 2 links i have estimate that the GTX1070 is around 2x faster than my GTX770 and the GTX1060 is around 1.8 faster than mine. That’s more logical than the blenchmark comparison among those 3 GPUS.

6

In your situation, I’d go with the Option 3 or 4. The GTX 1060 6GB

Thanks. For now I decided to get a GTX 1060. I’m still curious if a whole lot of processors would be competitive. It’s too bad memory is limited to the lowest GPU, you could load up with the lower-power 3GB cards and get quite a bit of horsepower for not a ton of money.

Somewhat related: I’m guessing hyperthreading doesn’t do much for rendering… as you really don’t get extra processing power out of the hyperthread. Is that true?

7

CPUs definitely can be competitive in certain complex scenes and I’m advocating to focus on CPUs for anything that’s professional work, simply because they’re more reliable.

However, from the options you outlined, I would pick the GPU option. Your CPUs are decent already and those 10-core Xeons have poor price-performance. Meanwhile, AMD is coming out with CPUs that undercut Intel’s offerings by almost 50%. The new professional (Opteron) lineup is yet to be known, but I’m sure AMD will have something appealing there as well. At this point in time, investing in CPUs would be unwise.

Somewhat related: I’m guessing hyperthreading doesn’t do much for rendering… as you really don’t get extra processing power out of the hyperthread. Is that true?

It actually makes quite a bit of difference on Cinebench, so I would expect it’s good for Cycles as well.

8

So… I got a GTX 1060 and added it to my machine… and after extended fiddling with the nvidia drivers, got everything up and running and decided to do some comparisons… in case anyone is interested in the results.

I picked a relatively simple scene to test with: Verts: 120,899 - 150 samples

First, pure CPU render: 8x 3.2Ghz
Tile size: 16x16
Speed: 1m 14.9s

CPU:
Tile size: 64x64
Speed: 1m 16.7s

Then I tried with just my original GTX-680 (also running my display)

GPU: GTX-680
Tile size: 256x256
Speed: 23.5s

Then I tried the new card by itself:

GPU: GTX-1060
Tile size: 256x256
Speed: 18.3s

Finally, I tried both together:

GPU: GTX-680 + GTX-1060
Tile size: 256x256
Speed: 11.6s

(I also did the GPU tests with a 512x512 tile size, but it was a bit slower in every case)

So, by my calculations my original GTX-680 renders about 3.2 times as fast as my 8 cores (non-HT) xeon @ 3.2Ghz.

Going from the GTX-680 to the GTX-1060, I get about a 30% improvement in render time, or 4.1 times as fast as my CPU.

Using both cards together, I get a full 2x speed increase over the 680 alone, or a full 6.4 times the CPU.

So… for the record: as far as bang for the buck goes, the GTX-1060 was clearly the right choice. I would have needed a full 52 CPU cores at roughly the same speed in order to equal the speed of the two GTX cards together.

Thanks again for the advice!