Ultra Shadow 2 = Full of it??

[bloodbob]

Where can you get the doco?
ftp:// : @torstenb.homeip.net/UltraShadow.pdf

Okay lets find the totally flaws in this POS.

In fact, the new technology in UltraShadow II allows for a 4× performance increase (compared to the previous generation) for passes involving shadow volumes— without the developer having to do any work.

Previous generation was of course UltraShadow 1 which gave a 2x performance increase against ATI cards ( this was b4 the 350 and 300 ) and older cards so thats a 8X fold increase over the R300 chip.

As a result UltraShadow II hardware can double the rendering horsepower to generate steshadow volumes at up to twice the standard pixel-processing rate.

This is what they state later well only off by a factor of 4?

Okay next lets go look at the hull extrusion?

Look at the original picture you will notice near the bottom next to the view weapon their is a shadow along the left hand side this would correspond to the player model shadow. Next look at the hull extrusion notice how their is no hulls passing through where the shadow was? umm we got shadows being magically created.

Now the major feature other the far and near clipping pertaining to bound boxes. Now a majority of the scene is all static level geometry of a bsp nature and their for it would mainly be triangle meshes. Now unless in the level creation tool it partition the meshes and assigns a bounding box for each then their probably should be no custom far or near clip for the worlds shadow ( other then of course the far/near render limits ). So in that unlikely case I could understand that their would be culling but I rather doubt this is the case.

Okay now if their really is culling going on the shadow volumes triangles would still need the point but you noticed the culling image triangles are simply cut in half and left with 2 sides this seems odd but could be some sort of rendering limitation in the drivers/engine.

Next if you actually carefully look at the 2 extrusion sets you will notice that they don't even have the same set of hull extrusions there are new ones in the culled image. If you look at the top of the image in the culled one near the middle but off to the right a little on the roof you will see a corner/vertex of a red box now if you go to the previous unculled image there is no red and its definitely not in the same place as any of the green hulls.

My whole impression is there isn't a single new feature in ultra shadow 2 and the images of the hull extrusions are fake.

Okay more gripes

Figure 6. Next-generation games use UltraShadow II to create photorealistic shadows, contributing to compelling digital experiences.

UltraShadow II is going to do little to nothing for increase the speed of soft shadows from any sort of none point light source.


Bloodbob.

Feel free to disagree with my point of view.

 

[DemoCoder]

NV3x could write 8 zixels per clock
NV40 can write 32 zixels per clock

32/8 = 4X

In non-multisampling, this seems to imply NV40 can write 8x as many zixels as R300.


The second quote seems correct to me. UltraShadow doubles the "standard" pixel processing rate. Standard = COLOR+Z fillrate.

 

[bloodbob]

NV3x could write 8 zixels per clock
NV40 can write 32 zixels per clock

32/8 = 4X

Thats great if you try to think of it that way but there is a subtle flaw which prevents that being true GF FX 6800 standard can't do 32 zixels per clock so either it doesn't have Ultrashadow II or they come about it by a different method ( such as picking numbers by random ).

 

[Mintmaster]

In non-multisampling, this seems to imply NV40 can write 8x as many zixels as R300.

Maybe I missed something, but doesn't R300 do 8 per clock too? That would make NV40 4x as fast, right?

 

[DemoCoder]

Yeah, brainfart on my part, I was up all night.

Something seems weird about what Dave's GT2 FSAA tests are showing.

On a non-ultrashadow card in z-write only mode, I would expect the no-FSAA stencil/zixel rate to be the same as the 2x MSAA stencil/zixel rate. Since, the extra z-units don't come into play unless you are in MSAA mode.

On an ultrashadow card (e.g. 8x0, or 32x0) I would expect the second ROP in the pipeline to be used instead to generate a second MSAA sample, so in 2xMSAA I would expect a 50% drop in stencil/zixel performance when MSAA is switched on.


But Dave's GT2 tests show that each level of MSAA only adds a 22% drop (e.g. 44% drop for 4xMSAA). There seem to be only a few explanations.

1) NV40 has the ability to 32x0 mode w/MSAA enabled possibly writing 3 Z values per clock (one ROP can generate 2, the other generate 1?)

2) Only about 30% of GT2 is stencil bound. The 50% performance drop only effects that 1/3. (e.g. let GT2 frame take X + .3 X to render, in 2xMSAA with 50% drop, it takes X + .6 X to render, .3x / (x + .3x) = 23%)

I wanna see a MSAA vs non-MSAA FableMark test.