Interesting topic. I worked a lot on the IFM performance in the last few days as it wasn't that good. I looked at the pCARS polylimits and tried to get somewhere near them. The problem was that we haven't had any numbers before I started on the car. So the polycount for the LODs aren't as low as I would want them to be. I would recommend four LOD levels: LOD1: up to 60,000 triangles LOD2: up to 25,000 triangles LOD3: up to 5,000 triangles LOD4: up to 1,000 triangles Those numbers include everything except driver and helmet model. It's also important to tweak your gen file and the LOD distances. I'm currently experimenting with those and I'm blending to LOD2 at 6 meters, to LOD3 at 18 meters and LOD4 at 35 meters. Polycount is really important in rF2. By optimizing the polycount I managed to gain over 20fps without a decrease in graphical quality.
I'd say, material properties are even more important. With model built with about 2-2.5 times more tris than Megane Trophy LOD0, but with just diffuse shader, I had more fps. But of course polycount optimization should always be performed and yeah, it's worth to spend more time on optimizing LOD distances too (but don't be too greedy on fps! ).
Well a ultra high poly model won't look great with a diffuse only shader Polylimits are really important and FPS are important as in the end this is a racing sim and not a interactive photo mode.
Yes, but it gives a good foundation for you to make it looking great... and also, because it is sim racing and not a look-a-like contest, I'd like to race in cars that look like their real life counterparts (with all the details), not something "almost/nearly as", like often seen in some older games But I think we can have nice models for first several meters and then shift to lower poly models (but still, without shifts beeing visible, as such transitions it doesn't look nice).
Well if you can't manage to create a model that looks like it's real life counterpart within 60,000 tris then it's more a problem of someones modeling skill and not the polylimit.
I have never saw such model made within that triangle limit. Of course, don't forget about detailed cockpit and other stuff, which also count to that limit.
I think this has always been very difficult problem. At least personally LOD switching kills my immersion much easily than simplified models or bad textures. I rather take mid detailed model which is seen from 0 to 500 meters than everything popping up all the time.
The pCARS models have a 55k tris limit for highest LOD (except for their ultra high detail stuff). That's without the cockpitview model. But they include everything else for the car (wheels, brakes, engine, suspension etc) except the driver. For the IFM mod the highest LOD has about 70k tris right now.
My ingame car has now over 150k on body alone, also quite lot of textures which many are not eved DDS, video ram is more of issue than actual performance with 8800GTS with only 0.5GB, I can have only two AI cars and I can't still use most of the tracks then, that much ram goes to models, with around 25k model I can have full grid, same textures. FPS are not huge problem really, sure it starts to slow down a bit but nothing huge, however running out of video ram is something one need to consider.
As soon as your driving in a full field it makes a huge difference. But it's not just the first LOD. I think the following LODs are even more important to get a good performance.
What I don't understand is why every model seem to take it's share of memory even they would be exact same cars, with textures I believe that same gets used to others models too, but with models that seems not to be the case, also memory consumption is so high that models seem not to have any compression when in gpu memory, but compressed DDS is in compressed form in memory, would be useful if models could have similar system. Then of course vertex damage etc requires models to be unique, maybe deformable causes them to be individual, haven't tested what happens there.
I guess every car needs an unique geometry because of the damage model. Also I never heard that models could be compressed in memory. I guess it would be an performance issue to decompress them on the fly.
I did read about DDS that it's performance benefit is that it does not need to be decompressed even it is compressed, which sounds odd to me, but it is marvel of modern technology. 3d models don't have that, I don't understand how it works with DDS but I would think that maybe in future they figure out how to achieve that with 3d models and can then use that to save memory usage. There are lot of objects on track that are never even touched, like many trees etc, in those there might be room for improvement in future when memory of gfx card is more limiting factor than performance.
Well for DDS: as far as I know it gets decompressed in realtime. One way to "save" memory usage and still get higher resolution meshes would be the tessellation feature of DX11. But again: 60k tris for your highest LOD (without a high def cockpit of course) is enough and you'll still be able to show the details on the cars. We should try to get as much out of that polycount before we should start talking about using 200k tris and up. A good modeler is able to create a very detailed car while still using as few triangles as possible. Anyone can create a good looking car while wasting triangles...
We would need to steal some modellers from first person shooters and alike, there just are too few of them :-\
I think we've got enough modelers in our community. We just have to be willing to improve so we can get the best out of a given polycount.