Graphics Card (GPU) based render engines such as Redhift3D, Octane or VRAY-RT have matured quite a bit over the last years and are starting to overtake CPU-based Render-Engines.
But what hardware gives the best-bang-for-the-buck and what do you have to keep in mind when building your GPU-Workstation? What is the best Hardware and best GPU for rendering with Octane, Redhsift3D or VRAY-RT, that is affordable?
Since GPU-Render Engines use the GPU to render, technically you should go for a max-core-clock CPU like the Intel i7 7700K that clocks at 4,2GHz (4,5Ghz Turbo) or the Intel i7 8700K that clocks at 3,7Ghz (4,7Ghz Turbo).
That said though, there is another factor to consider when choosing a CPU: PCIE-Lanes.
GPUs are attached to the CPU via PCIE-Lanes on the motherboard. Different CPUs support different amounts of PCIE-Lanes and Top-tier GPUs usually need 16x PCIE 3.0 Lanes to run at full performance.
The i7 7700K/8700K has 16 PCIE-Lanes meaning you could use only one GPU at full speed with these type of CPUs. If you want more than one GPU at full speed you would need a different CPU that supports more PCIE-Lanes like the AMD Threadripper CPUs, that have 64 PCIE-Lanes, the i9 7800X (28 PCIE-Lanes) or the i9 7900X Series CPUs that support 44 PCIE-Lanes.
GPUs, though, can also run in lower speed modes such as 8x PCIE 3.0 Speeds and then also use up less PCIE-Lanes (8x). Usually there is a negligible difference in Rendering Speed when having GPUs run in 8x mode instead of 16x mode.
This would mean you could run 2x GPUs on an i7 8700K in 8x PCIE mode, 3x GPUs on an i9 7800X and 5x GPUs on an i9 7900X. (Given the Mainboard supports this configuration)
When actively rendering and your scene has been loaded into the GPU VRAM, it fits nicely in there and nothing has to be swapped out of core, GPU renderers are of course mainly dependent on GPU performance.
Some processes though that happen before and during rendering rely heavily on the performance of the CPU, Hard-Drive and network.
For example extracting and preparing Mesh Data to be used by the GPU, loading textures from your Hard-Drive and preparing the scene data.
In very complex scenes, these processing stages will take lots of time and can bottleneck the overall rendering speed, if a low-end CPU, Disk and RAM are employed.
If your scene is too large for your GPU memory, the GPU renderer will need to access your System RAM or even swap to disk, which will considerably slow down the rendering.
Best Graphics Card for Rendering
To use Octane and Redshift you will need a GPU that has CUDA-Cores, meaning you will need a NVIDIA GPU. VRAY-RT additionally supports openCL meaning you could use an AMD card here.
The best bang-for-the-buck NVIDIA cards are 1070 GTX (1920 Cuda Cores, 8GB VRAM), 1080 GTX (2560 Cuda Cores, 8GB VRAM) and the 1080 Ti (3584 Cuda Cores, 11GB VRAM).
On the high-end, the currently highest possible performance is offered by the NVIDIA Quadro P6000, that also comes with 24GB of Video RAM. This Card though has horrible Performance per Dollar (Check the tables further down).
NVIDIAs new Volta architecture, that is being released for consumers soon, will top this performance even more. Note though, that the Render Engines you are using will have to be updated in Order to use the New CUDA 9 Architecture these Volta Cards run on.
GPUs, that have 12GB Video RAM and more, can handle high-poly scenes with over 200 million unique objects best. Take a look at the performance per dollar tables below, though, to get an overview of how expensive some of these cards can get without offering that much more performance.
Founders Edition Blower Style Cooler
- PRO: Better Cooling when stacking more than one card
- CON: Louder than Custom Partner Card Cooling
Custom Partner Cooling
- PRO: Quieter than Blower Style, Cheaper
- CON: Worse Cooling when stacking cards
- PRO: Best All-In-One Cooling for stacking cards
- CON: More Expensive, needs room for radiators in Case
- PRO: Best temps when stacking cards, Quiet
- CON: Needs lots of extra room in the case for tank and radiators, More Expensive
Be sure to get a strong enough Power supply for your system. Most Cards have a TDP of around 180-250W. CPU of around 100W and any additional Hardware in your case.
I Recommend a 500W for a One-GPU-Build. Add 250W for every additional GPU. Good PSU manufacturers to look out for are beQuiet, Seasonic and Coolermaster (among others)
Mainboard & PCIE-Lanes
Make sure the Mainboard has the desired amount of PCIE-Lanes and does not share Lanes with Sata or M.2 slots. Also be careful what PCI-E Configurations the Motherboard supports. Some have 3 or 4 PCI-E Slots, but only support one x16 PCI-E Card.
This can get quite confusing. Check the Motherboard manufacturers Website to be sure the Card configuration you are aiming for is supported. Here is what you should be looking for in the Motherboard specifications:
In the above example you would be able to use (with a 40 pcie Lane CPU) 1 GPU in x16 mode. OR 2 GPUs in both x16 mode OR 3 GPUs one in x16 mode and two of those in x8 mode and so on. Beware that 28 pcie Lanes CPUs in this example would support different GPU configurations than the 40 lane CPU.
Currently the AMD Threadripper CPUs will give you 64 PCIE Lanes to hook your GPUs up to, if you want more you will have to go the multi-CPU route with Intel Xeons.
To confuse things a bit more, some Mainboards do offer four x16 GPUs (needs 64 PCIE-Lanes) on CPUs with only 44 PCIELanes. How is this even possible?
Enter PLX Chips. On some motherboards these chips serve as a type of switch, managing your PCIE-Lanes and leads the CPU to believe fewer Lanes are being used. This way, you can use e.g. 32 PCIE-Lanes with a 16 PCIE-Lane CPU or 64 PCIE-Lanes on a 44 Lane CPU. Beware though, only a few Motherboards have these PLX Chips. The Asus WS X299 Sage is one of them, allowing up to 7 GPUs to be used at 8x speed with a 44 Lane CPU, or even 4 x16 GPUs on a 44 Lanes CPU.
This screenshot of the Asus WS X299 Sage Manual clearly states what type of GPU-Configurations are supported (Always check the manual before buying expensive stuff):
PCIE-Lane Conclusion: For Multi-Gpu Setups, having a CPU with lots of PCIE-Lanes is important, unless you have a Mainboard that comes with PLX chips. Having GPUs run in x8 Mode instead of x16, will only marginally slow down the performance. (Note though, the PLX Chips won’t increase your GPU bandwith to the CPU, just make it possible to have more cards run in higher modes)
Best GPU Performance / Dollar
Ok so here it is. The Lists everyone should be looking at when choosing the right GPU to buy. The best performing GPU per Dollar!
This List is based off of OctaneBench 3.
(It’s quite difficult to get an average Price for some of these cards, since crypto-currency mining is so popular right now, so I used MSRP)
|GPU Name||VRAM||OctaneBench||Price $ MSRP||Performance/Dollar|
|GTX 1070 TI||8||130||450||0,288|
|GTX 1080 TI||11||186||700||0,265|
|GTX TITAN Z||12||144||2999||0,048|
And here is a List based off of Redshift Bench, note how the cards scale (1080TI) [RedshiftBench Mark (Time [min], shorter is better)]:
|GPU Name||VRAM||RedshiftBench||Price $ MSRP||Performance/Dollar|
|GTX 1080 TI||11||11.44||700||1,248|
|4x GTX 1080 TI||11||3.07||2800||1,163|
|2x GTX 1080 TI||11||6.15||1400||1,161|
|8x GTX 1080 TI||11||1.57||5600||1,137|
Source: Redshift Forum
And here is a List based off of VRAY-RT Bench. Note how the GTX 1080 interestingly seems to perform worse than the GTX 1070 in this benchmark:
|GPU Name||VRAM||VRAY-Bench||Price $ MSRP||Performance/Dollar|
|GTX 1070||8||1:25 min||400||2,941|
|GTX 1080 TI||11||1:00 min||700||2,380|
|2x GTX 1080 TI||11||0:32 min||1400||2,232|
|GTX 1080||8||1:27 min||550||2,089|
|4x GTX 1080 TI||11||0:19 min||2800||1,879|
|TITAN XP||12||0:53 min||1300||1,451|
|8x GTX 1080 TI||11||0:16 min||5600||1,116|
|TITAN V||12||0:41 min||3000||0,813|
|Quadro P6000||24||1:04 min||3849||0,405|
Source: VRAY Benchmark List
Speed up your Multi-GPU Rendertimes
So unfortunately GPUs don’t scale linearly. 2 GPUs render an Image about 1,8 times faster. Having 4 GPUs will only render about 3x faster. This is quite a bummer, isn’t it? Having multiple GPUs communicate with each other to render the same task, costs so much performance, that one GPU in a 4-GPU rig is basically just managing decisions.
The solution is the following: When final rendering image sequences, use as few GPUs as possible per task. If you have to render 4 images and have 4 GPUs, let every GPU render one image instead of having 4 GPUs render on every image. This way you will not encounter any slow-down. Some 3D-Software might have this feature built-in, if not, it is best to use some kind of Render Manager, such as Thinkbox Deadline (Free for up to 2 Nodes/Computers).
Beware though, that you might have to increase your System RAM a bit and have a strong CPU, since every GPU-Task needs its amount of RAM and CPU performance.
What Hardware configurations are you eyeballing with, or have already been able to test? Share your thoughts in the comments!