0
$\begingroup$

A few days ago I finished a build, including two Titan XP. The goal was to get faster rendering, letting the gpus work as two seperate cores that render two chunks at the same to nearly double the rendering-performance. So I disabled SLI in Nvidia Control Panel, selected both gpu's in the preferences tab and selected Cycles, using a simple scene at 4k.

The performance of the Titan XP should be circa at the level of the 2080 TI, as I saw in a few benchmarks. So two of them should be significantly faster (I read that two identical gpu's, that are not running in SLI should nearly double the performance). But they rendered nearly at the same speed as the 2080 TI/a single Titan XP, so there was no improvement. CPU, Mainboard and RAM shouldn't bottleneck (7980XE, 32Gb, Asrock Taichi CLX), I'm using Blender 3.0 with no add-ons and I installed the newest graphic drivers.

I tested a few different things: only selecting one of the cards in Blender - both of them work perfectly fine

monitoring the load of the gpus - both run at 100% with SLI disabled (MSI Afterburner)

removing the SLI-bridge - changed nothing

enabling SLI - was a little bit faster than a single card, but maybe only about 20-30%

trying different tile/chunk sizes

But nothing helped.

So what can I do, to get the cards work together? What have I forgotten? And is it possible to get the cards working on different chunks/tiles in Blender 3.0 as I described above?

$\endgroup$

1 Answer 1

1
$\begingroup$

As explained on rendering - No Auto Tile Size in Blender 3.0 - Blender Stack Exchange: Blender 3.0 Cycles rendering works significantly different from previous versions. Especially when it comes to tile rendering. Using tiles to force your GPUs to work on different areas might not be the way to go anymore, I'd suggest you to try without that.

Basically do not follow any trick or advice that was specific to Cycles prior 3.0 because they are effectively fit for a different render technology.
Instead, seek out advices for "Blender 3.0" and up, or "Cycles X".

Also:

I read that two identical gpu's, that are not running in SLI should nearly double the performance

That is kind of untrue. Adding X amount of identical GPU doesn't necessarily divide your render times by X. That is a "best case scenario", which you don't reach all the time. It does make things faster, but it's not as straightforward and it also depends the scene.

$\endgroup$
1
  • $\begingroup$ Thank you very much, I think that I searched in the wrong direction and just had a too simple render. So I chose a more complex render (Spring): Single Titan: 2:42 Dual Titan: 2:01 It was an improvement compared to the single Titan, so the dual took about 75% of the time of the single card, so in this case I got a good performance boost. And maybe I will get a bigger boost, rendering a more complex scene. Doubling the resolution made the performance gap bigger: Single Titan: 7:06 Dual Titan: 4:11 So it took only 58% (nearly 200% of the performance) of the time. $\endgroup$
    – Bone 531
    Jan 7, 2022 at 21:19

You must log in to answer this question.

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