Physics Performance MeshImpostor vs multiple PrimitiveImpostors

I have some objects which need collision, that can’t be modelled with single primitiveImpostor. For now I’m using the mesh Impostor for that. But I wonder if there is a roule of thumb on when one should go for multiple primitives vs one mesh. Has anyone done some testing on that or are there ressources available?

In my exact case it is 6 Boxes vs a Mesh for a six sided tube.

In case someone has the same question:

I did some testing myself with ammo.js and got quite interesting results.
According to babylon stats my scene has around 39,000 vertices in total.
Once I divided my scene into 330 primitive colliders and once I merged everything into 1 big mesh (except the player). I tested it on an older I7-quadcore with 2.5ghz and integrated intel graphicscard.

With the 330 primitive colliders i got a pretty stable 2-5 fps, with Animation duration per frame at 100-160ms.
With the whole scene as a single mesh I had solid 60 fps, sometimes dropping to 40. Animation duration was often under 1ms here.

Has anyone an Idea how why the single mesh collider is so much faster even though the docs state how inefficient they are?

Pinging @trevordev

But I think this is an order of magnitude. One sphère impostor vs one mesh impostor will obviously see a win of the sphere… But 330 spheres impostors vs 1 mesh could see the opposite as having so many impostors can clearly be slow

Yep, David’s answer looks correct to me. I would try to test compound primitive impostors vs a single mesh impostor. I would expect the compound primitive shape to be more performant than a complex mesh impostor.