1
$\begingroup$

enter image description here

Hi,

I've been using Blender 2.78a on Windows 10 Pro. The problems I'm trying to solve and I search for many solutions around the www is the classical Memory CUDA problems. I use a GTX 970 and batch render as you can see on the image. It has 4 Gb VRam and I use less than 2 Gb for the render... Seems Cycles can generate the first bucket but when it goes to the second always crashes. I don't understand anything and it's so frustrating for me... Things I've done: - http://artificialflight.org/blog/2013/cycles-crash-cuda-tdr-error/ I put 64, 16, 256... - I have last drivers - Reduce bucket size

Could anyone help please? Thanks in advance

$\endgroup$
1
  • $\begingroup$ Why are you using an old version of Blender? $\endgroup$
    – dr. Sybren
    Jun 18, 2017 at 14:40

2 Answers 2

0
$\begingroup$

Searching a little bit more and more, I find many things about Mem Cycles consume and real consume. With GPU-Z I can see a real read. First of all, the modifiers turned on in render, Subdiv mainly. When blender Mem says 1670Mb, is the limit for my 4Gb GTX 970. All textures is better to be JPG, PNG. Only use float as EXR or HDRI on particular cases like environment lighting.

Maybe, one day, Cycles developers would think using GPU only to process and hardware RAM to cache everything, like Redshift and Octane do :)

$\endgroup$
3
  • $\begingroup$ Yeah, I don't get why the GPU doesn't just read from the RAM! Why is that? Maybe the reading would be a little slower, but I bet the processing capability would increase so much it would be worth it. Or maybe it could be a checkbox option. That's a great idea! $\endgroup$ Jan 3, 2017 at 19:56
  • 1
    $\begingroup$ I think it's the key for not to be left behind and let anyone with a more or less good gpu to render with it. If Redshift and Octane are doing this is because it's not a whim and it's something possible :) $\endgroup$ Jan 4, 2017 at 18:17
  • $\begingroup$ Yeah, I agree. Blender's great but there are always ways for it to be getting better. $\endgroup$ Jan 4, 2017 at 19:24
0
$\begingroup$

Here ir the problem of exceding the memory on CUDA:

On sampling: AA = 4 Diffuse, Reflect... = 128 It makes a memory bug where run out of

But: AA = 128 Diffuse, Reflect... = 4 Goes great and the vram is under control

Maybe maths are the same, but not the result. My conclusion is to work always first with AA and put other samples at 0. When you have a good result you can start tweeking the other samples to clean the noise.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .