4
$\begingroup$

I'm going to build my new workstation/rendermachine and im stuck with the question: "how much vram do i actually need?"

More gpu accelerate the rendering process, but the max vram load is given by the individual vram amount. So im torn between 4x8GB Vram and a single 32GB vram gpu.

1) What is the mathematical context between verticies/faces and vram load?

2) Could someone show me a rendered scene or video of how much I could do with 32gb of vram.

Thanks alot

$\endgroup$
1
  • 2
    $\begingroup$ I've got a 6Gb GTX980ti card and never hit that limit rendering mostly architectural scenes. Even my 1.5Gb has enough memory for 95% of what I do. Generally large textures are the biggest memory hogs during rendering, geometry is rarely an issue. Since you never state what type of work to do it is hard to recommend anything, but I'd say 4x8Gb would give you far better performance and more that enough memory. $\endgroup$ Commented Apr 16, 2018 at 22:33

2 Answers 2

7
$\begingroup$

I bet those 4x8GB gpus are much faster than single 32GB card. Here are some options:

  • Get the faster setup and a renderer that can do out-of-core rendering like Octane or Redshift. This allows your GPUs to use system RAM while rendering in exchange of slightly lower render speed. This is something Cycles can't do yet. If you are spending money on hw you might as well get a software.

  • Get the faster setup and if you hit VRAM limits optimize your scene - you can split the rendering into multiple render layers which you would be probably doing anyway on such large projects.

  • Getting the 32GB card has also benefits. This comes handy when doing physics simulations accelerated by openCL and you want to fit the ocean or the whole smoke explosion into your card. Blender is not there yet regarding simulations and can't accelerate it on gpu, but if you were using Houdini or something similar you might benefit and need this much VRAM. Simulations can't be easily split like rendering can.

8GB of VRAM is quite a lot and definitely sufficient even for serious production use. Most will be taken by:

  • ~1GB for every 8 million triangles
  • 500MB for Cycles kernel
  • 64MB for each 4k texture (8bit x RGBA)
  • 200MB for each 4k HDR texture (32bit per channel)

When you are not rendering you might want to mine crypto on those gpus and maybe those 4x8GB cards will pay themselves faster than that single 32GB one. Another thing to consider maybe.

$\endgroup$
1
  • $\begingroup$ That was a quick and detailed answer with everything i could wish for. $\endgroup$
    – 0to100
    Commented Apr 17, 2018 at 0:19
0
$\begingroup$

8M triangles taking 1GB of VRAM is the theory, but if you open blender, cycles viewport, makes the default cube 8M tris, you see blender taking 6GB of Vram.. And if you take a terrain or a plane, and adaptive subd with a few viewport subd, until just 500K tris, it also takes 6GB of Vram. I don't know why, but Vram gets filled explosively with simple stuff. If the plane is subd normally (no adaptative) until 32M polys, 4x8M bascally, it does NOT take 4GB of VRAM, it takes 16GB !

That's crazy but basically default cube to 8M tris adds 16GB of VRAM, NOT 1GB, for viewport cycles.

In EEVEE & workbench, 8M DOES equals 1GB vram though.

Also crazy, if two viewports are active, it doubles the VRAM usage, like if it stored the whole scene twice in VRAM. So with only 1 cycles viewport 8M tris takes 8GB of Vram, 16GB with twos. (I usually work with two....) But not so much with default cube, or maybe because it doesn't flush objects that have been deleted and purged from the blend file, they stay in VRAM.. Really confusing. And when rending, all that data will be copied again to the VRAM on top of the viewport ones, but takes way les space, 10GB of viewport VRAM only transcribe as 2GB for render, for exactly the same subdivs/tris count of 8M.

Also, denoise does take a flat 3GB fee on the vram, bam. on/off simply adds/remove 3GB instantly.

So bottom line, with 8GB you can subdivide a cube 8 times, and that's it.

Just as a visual proof:

poly count:

poly count

total VRAM:

total vram

blender only VRAM:

blender only vram

Vram used for rendering, probably flushed when taking all this screenshots:

cyles render

The 8M poly cube:

cube subd 9 times

subd 9 times:

9 subds

$\endgroup$

You must log in to answer this question.

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