Indigo Renderer Forum

The old Radeon 4xxx series were never meant for general-purpose compute, and aren't even enough to run the existing Indigo GPU acceleration.

But in any case, you can get quite cheap and powerful Radeons these days; they use a lot of power and generate a lot of heat, though...

Thanks lyceum, for answering my earlier questions.

I have a gtx 970, When you get one to work with I'm curious if the 3.5gb (fast) + .5gb (slower) memory division causes any issues with the pure gpu render mode? I've done some distributed rendering on it with the current stable indigo build and haven't seen any issues yet.

If you want more info about the 970gtx memory split before getting a card to test anandtech's article is a good place to start: http://anandtech.com/show/8935/geforce- ... allocation

Also, I and probably half the forum users here would love to help run tests if you're looking
My best workstation is a water cooled (but not OC'ed) i7 5960x+16gb mem+Evga GTX 970+windows 7 pro (only a render node license is on it at present).

I've followed the controversy of course, and in my opinion it's both massively overblown, and people aren't even noticing what I think is the bigger issue: the missing cache due to the disabled units in the 970. It is exactly this extra cache memory that makes a lot of the performance improvement in Maxwell over previous Kepler generation, and similarly is responsible for the Intel iGPUs being so strong.

Here are some OpenCL benchmarks on TechReport showing how much cache helps on eg LuxRender: http://techreport.com/review/24879/inte ... eviewed/13

Bottom line from my point of view: Sure, there will be some situations where 3.5GB is not enough, but in what fraction of those cases is even the full 4GB not enough? Probably that extra 500mb is not going to be make-or-break every single time, and if it is, then maybe you should rather get an 8GB Radeon for cheap, or a GTX 980 Ti with 8GB (coming soon).

So will Indigo suddenly perform much slower if you scene takes more than 3,5 GB of VRAM?

Dunno for sure, haven't tested with a 970.

On the other hand, Nvidia make a perfectly fair point when they say (paraphrased): it could be worse, you could not have the extra 500MB at all, which would perform far worse (due to PCI Express swapping). Also, it's hugely discounted compared to the full 980, so I don't see how people can get so upset about it.

Those complaining about the 500MB should rather worry about the missing cache, that the memory bus is only 256bit on such an expensive GPU, that we're still on 28nm after who knows how long... things like this that actually have a big influence

The insight is appreciated. The article too, at least as much of it as I understood.

As you mentioned I don't think anything I've read on it more than mentions the lost cache. I didn't realize it played such a large role in compute. Then again, intel's been expanding cache on all their chips for years...exercising that philosophy in the iris pro - I've heard such decisions are not taken lightly due to added cost and size.

I'd probably still would have still bought gtx970 had I known in advance as I'm working on learning 3ds max and it seems to favor NVidia cards and it fit my budget. I'm just annoyed finding out after I bought it. Maybe this years upgrade budget will go toward a Titan X and an indigo upgrade when 4 is released .

lycium wrote: The only way you're going to have more memory and rendering power is with a Titan X, which is ridiculously expensive. Ask kklors

Haha just seen this one. After all, 12GB is what gets me into GPU rendering in first place : P

We talked about optimizations and so on before, but sometimes it's really not the goal or appropriate for a scene that has to be done asap and will never be opened up again spending man hours to optimize rather than creating content. One could create LODs for trillions of levels on geo and textures, create cam facing billboards for far away objects etc., but that's sort of UE4 land nowadays : D

So far I had 3 GPU render gigs and two of them were a major pain in the ass to squeeze into the VRAM.
In both cases rendering on GPU felt highly counter productive (very objectively in depth), so 12 is at least getting somewhat close.

How are those benchmarks coming along then?

Are the benchmarks referring to me? Even if not; I want them! : D

No, lycium said he was going to run benchmarks last week to compare GPU's. Would be interested to see the results.

It's turning into a test suite for GPU rendering on all platforms, which can then stunt double as a benchmark Sorry for the delay!

On the plus side, I have the opportunity to test on some really high end hardware again soon, so should be worthwhile

I am rather excited now to see some benchmarks. I just built my computer the other day and am only rendering with my i7 5820k for now - and the results were already alarming compared to my i3 at work.

I just hope I can wait to buy a new GPU

This wait has been so long I can't even remember my life before it.

The "benchmark" is more of a test suite at the moment, used for comparing the output of the GPU renderer to the CPU reference images for a bunch of scenes and GPU platforms.

Another reason it's taking longer than expected is that many of our "nice" test scenes have complex materials (eg ISL shaders) which are not supported by GPU yet. On that note, if anyone has some nice scenes with relatively simple materials (and ideally without specular) I'd be really happy to try them out!