I haven't quite got around to try this yet but I was thinking it over and was thinking it may prove a problem.

My rendering framework so far draws any opaque meshes first, the Z-order isn't considered important what with the depth buffer. After that "cheap" transparent frames are drawn, also in no particular order but at least after the opaque ones and finally I sort the remaining transparent frames by their screen depth so that a transparent object behind another transparent object doesn't get clipped by being drawn after the first one.

However if I'm to use instancing for transparent frames that would benefit from the latter approach... this would be a problem wouldn't it (since they're all rendered in the same draw call and as such can't be sorted in with other meshes)? Is there any good solution to this or am I missing some point to instancing which means this really isn't an issue?

I presume this means additive or screendoor transparency effects?

Not really. As I've been lead to understand by various articles changing rasterizer states, blend states and shaders were to be considered more of a bottleneck (in general) than things like draw order. I suppose it makes sense that it can be the other way around if you have a sufficiently complex set of shaders going though; at the moment I don't though so such optimizations haven't really crossed my mind yet. So basically when I said "cheap" transparency I mean they are just drawn after the completely opaque meshes but prioritize state changes over depth when it comes to sorting. This will likely be good enough for things like floor covers. On the other hand if you have a window and can see a (semi-transparent) body of water through it it wouldn't do if the window was rendered first and alpha-blended with the backgrond before the water surface had been written there, so in that case it makes more sense to prioritize depth over state changes.

Just sort the instance transform data back-to-front before rendering and you'll be fine.

I thought about that but then if the instances are scattered about a great deal, wouldn't non-instanced (or other instances for that matter) meshes that are positioned in-between fail to end up where they should in a depth sort?

If you want correct output, you _must_ draw things in the necessary order, whether it's a bottleneck or not.

I'm aware of this, just to make it clear when I was talking about cheating with "cheap" transparent objects I didn't really mean translucent objects but rather close to 1-bit transparent things like floor covers or tree leaves. There may be some blending but it only occurs along the very egdes then and it therefore can be argued it isn't going to be noticible whether these edges blend properly with other transparent objects behind them. Those kinds of meshes would likely make up most of my transparent ones and actually semi transparent ones won't be very many.

So what I then understand is that you can (and should) split up instances in several batches when rendering?

I didn't know you could do that (as said I haven't actually gotten around to implement this yet, will see if I get the time during the weekend) but the approach outlined by AgentC seems to make sense then.

I'm sure I'll be back with more issues before long, but until then thanks for the descriptions and hints.