Indigo Renderer Forum

Now that we got a sneak peek at what full GPU rendering in Indigo will look like, I think many of us are starting to considering upgrading to better graphics. Question is - what graphics cards provide the best bang for the buck in OpenCl rendering? I've heard nVidia's 500 series is supposed to be better than the generation after it for example.

Also interested to know if my present GPU, planned especially for GPU rendering, was a good choice.

Min VRAM? Min OpenCL version? Best bus size?

Don't speculate about unfinished software!
Why not wait until a public Beta is released and Benchmarks are available?
There is no sense in buying Hardware now, it may be outdated once it can be used

I'm not speculating - I'm asking the developers. You know, those who knows the ins and outs of the software?

I need a workstation now - not whenever the public beta is released. Thought I might as well get some good graphics while I'm at it.

Oscar J wrote:I'm not speculating - I'm asking the developers. You know, those who knows the ins and outs of the software?

Afaik the devs don't have that many Rigs with many GPUs, therefor I pointed to some community made benchmarks once a beta is out. But you are right they maybe can hint you in the "right direction"...

Oscar J wrote:I need a workstation now - not whenever the public beta is released. Thought I might as well get some good graphics while I'm at it.

in that case go for it

best bang for buck would probably be an AMD R9 290x with 8GB of VRAM. Id assume the the GPU rendering slows considerably when data has to flush back and forth between system and gpu RAM, also the general OpenCL performance of a 290x is top notch.

Personally I'm looking at nVidia cards because they work much better with blender.

then do it up but id make sure you get a high vram card or else all that speed isnt going to do a whole lot. buying one card that lasts a long time is much cheaper than hitting your ram limit and having to buy another card all over again. I cant say for sure as I dont have the software but I would imagine an 8GB card is going to be leaps and bounds better for this type of application. and if all your materials can fit directly on the vram then id imagine performance would be blistering. More VRAM in theory would render more complex scenes faster.

AFAIK the AMD cards outperform the nVidia cards in the OpenCL department

Juju wrote:AFAIK the AMD cards outperform the nVidia cards in the OpenCL department

It depends on the application. For Chaotica I've found this to be true (smaller, simpler kernels), but for Indigo (very complex path tracing kernel) NVIDIA's new Maxwell architecture is really impressive.

I'm borrowing a GeForce GTX 980 from a friend at the moment, and it's amazingly fast. Our own GeForce GTX 750 Ti (first generation Maxwell) performs like a high end GPU, despite being only 100 Euros or so.

As fourzeronine said, 8GB memory on GPUs will be fantastic for GPU rendering. However, we're finding that we can only allocate 25% of available memory on NVIDIA GPUs using OpenCL, which is really annoying and practically limiting. AMD GPUs allow you to use quite a reasonable %... so there seems to be some kind of tradeoff here unfortunately

In short, for best rendering performance, I'd recommend NVIDIA GPUs at the moment, obviously with as much memory as you can get. However, currently it's going to run into memory limitations quite easily - we'll contact NVIDIA about this soon and see if there's anything that can be done about it.

lycium wrote:...............

In short, for best rendering performance, I'd recommend NVIDIA GPUs at the moment, obviously with as much memory as you can get. However, currently it's going to run into memory limitations quite easily - we'll contact NVIDIA about this soon and see if there's anything that can be done about it.

This is quite interesting since in all OpenCL benchmark I did look AMD was outperforming Nvidia, except Compubench where strangely 980 is stated on OpenCL 1.2.
Nvidia is finally enabling OpenCL 1.2 on drivers?
What's the max VRAM allocation you experienced so far?

We still don't have OpenCL 1.2 on NVIDIA, and it's majorly annoying! You can't even clear a memory buffer without writing and compiling a little kernel for it...

As I mentioned the max mem alloc I'm seeing on NVIDIA is 25%, i.e. 1GB on a 4GB card. Attached are some screenshots with stats, unfortunately I don't have our AMD FirePro W5000 connected at the moment due to the PSU only barely managing the 980's power requirements, but if I recall correctly it allowed 1.37GB out of 2GB to be allocated.

I should add that I've tried various things I found online, setting environment variables for max allocation % etc, all are happily ignored by the current drivers.

How much memory can the 750 Ti allocate? Is it still a fourth even though it's got less memory?

Also - what's the ks/s difference between the 750 ti and 980?

Unfortunately yes.

I probably shouldn't say this, but I'll bet this is deliberate thing to push people to CUDA instead of OpenCL.