A
AirRaid Mach 2.5
http://www.beyond3d.com/forum/showthread.php?t=37033
QUOTE:
"I'm a programmer I assure you, been one for 12 years in the games
industry. No, I can't say my employers name because my comments could
be construed as representing their opinion, which would get me into
trouble."
"There are *very* few instances where one would pick the PS3 gpu over
the 360's because the PS3's gpu is weak compared to 360's. That's not
just my opinion, ask other devs how they feel about it. Sony chose
poorly when they chose the video hardware. Performance wise, the 360's
video will out render the PS3's everytime if you feed it the same 3d
scene. The only way to make the PS3 keep pace is to leverage its spu's
to preprocess your 3d scene."
"I have a shader that needs 8 vertex inputs, Position, Normal, Color,
and Texcoord1 thru Texcoord5. Assume all inputs are packed (ie, .x, .y,
..z and .w are all used). Further, I need this shader to be applied to a
mass of 100 enemies that are all close to the camera (and hence using
their best lod) and are 5000 verticies each. Or if you prefer, imagine
that it needs to be used on a 3d scene of 500,000 verts. These
scenarios are quite common, and they *kill* RSX performance. If you
don't believe me, just read the RSX docs, its right in there and it
literally says "this kills RSX performance". The only way to
accommodate these scenarios is to precull using the spu's. If you just
rely on the RSX to do it for you, then you will never match the xenon's
framerate, period. This isn't conjecture or opinion, its hardware
design fact. Again, I encourage you to read the RSX docs as well as PS3
dev forums, I'm not the only one complaining about this."
"I'm a 360/PS3 programmer by trade, focusing on graphics and
optimization. It's fairly rare that I post, but I just got off a long
crunch and have been browsing the forums more than normal to unwind
Incidentally most of that crunch was spent trying to get the PS3 to
match the 360's frame rate. It falls short, but it hits 30fps so we're
ok.
The Gran Turismo image posted above is a good shot, but its a testament
to the talent of the dev crew, not just the hardware. Vertex/pixel
shaders are written in HLSL and can be compiled to run on either
PS3/360 although you can tweak them to suit the console. I guess my
point here is that if you take that same vehicle 3d model and those
same shaders that make it look all pretty, and compile/run them on the
360, it will look the same and run at a higher framerate than on the
PS3. This is because the PS3's gpu isn't as good as the 360's."
"I'm assuming you also know that the PS3's vertex processing units are
terrible, since each extra vertex shader input ads one cycle of delay.
Likewise, you probably also know that the only way currently around
this limitation on PS3 is to use the spu's to preprocess all geometry
by backface culling them first on cell before feeding them to the gpu.
But then you are still stuck with other PS3 gpu limitations, such as
not being able to do anti aliasing with floating point render targets
so you can't have MSAA and HDR simultaneously."
"The vertex pipeline doesn't matter when you have to waste one cycle
per vertex input. In other words, the gpu stalls untill it can fetch
all the data it needs before it can even start executing the vertex
shader. Why is this important? Because next gen games needs lots of
lookup maps to look nice, which means lots of u/v coordinates and other
data that needs to be passed to the vertex shader, ie, lots of inputs.
In RSX's case, that means gpu stalls. This is the RSX's well known
achiles heel and is well documented. The only known work around at the
moment is to use the cell spu's as another 'gpu', in this case a
culling gpu, to limit the number of verts actually sent to the RSX.
Whether or not RSX+SPU culling will help PS3 meet or exceed xenon's gpu
performance isn't known at this point."
"You don't think no msaa with floating point buffers isn't a huge
limitation? There are other color spaces, but they are useless to me. I
don't need 8/8/8/8 int. I need 8/8/8/8 float. In RSX's case you have no
choice but to use FP16 (16/16/16/16), compared to FP8 on xenon. So you
are forced to move around twice the memory on RSX if you want a
floating point buffer, which means less framerate. Just try it! I'll
assume your a game dev. Switch your PS3 game from FP16 to an 8/8/8/8
int format and see your framerate jump. Of course, you'll have to forgo
HDR on your shipping title, but you can then do msaa. Or, go back to
FP16 since HDR looks so cool, but oh ya, you then have to turn off
msaa. I just don't have these headaches on 360, but I have to deal with
them PS3."
____________________________________________________________
sounds like he knows what he is talking about and is not bullshitting.
everything he is saying fits well with what other developers have been
saying for many months, even a year or more.
QUOTE:
"I'm a programmer I assure you, been one for 12 years in the games
industry. No, I can't say my employers name because my comments could
be construed as representing their opinion, which would get me into
trouble."
"There are *very* few instances where one would pick the PS3 gpu over
the 360's because the PS3's gpu is weak compared to 360's. That's not
just my opinion, ask other devs how they feel about it. Sony chose
poorly when they chose the video hardware. Performance wise, the 360's
video will out render the PS3's everytime if you feed it the same 3d
scene. The only way to make the PS3 keep pace is to leverage its spu's
to preprocess your 3d scene."
"I have a shader that needs 8 vertex inputs, Position, Normal, Color,
and Texcoord1 thru Texcoord5. Assume all inputs are packed (ie, .x, .y,
..z and .w are all used). Further, I need this shader to be applied to a
mass of 100 enemies that are all close to the camera (and hence using
their best lod) and are 5000 verticies each. Or if you prefer, imagine
that it needs to be used on a 3d scene of 500,000 verts. These
scenarios are quite common, and they *kill* RSX performance. If you
don't believe me, just read the RSX docs, its right in there and it
literally says "this kills RSX performance". The only way to
accommodate these scenarios is to precull using the spu's. If you just
rely on the RSX to do it for you, then you will never match the xenon's
framerate, period. This isn't conjecture or opinion, its hardware
design fact. Again, I encourage you to read the RSX docs as well as PS3
dev forums, I'm not the only one complaining about this."
"I'm a 360/PS3 programmer by trade, focusing on graphics and
optimization. It's fairly rare that I post, but I just got off a long
crunch and have been browsing the forums more than normal to unwind
Incidentally most of that crunch was spent trying to get the PS3 to
match the 360's frame rate. It falls short, but it hits 30fps so we're
ok.
The Gran Turismo image posted above is a good shot, but its a testament
to the talent of the dev crew, not just the hardware. Vertex/pixel
shaders are written in HLSL and can be compiled to run on either
PS3/360 although you can tweak them to suit the console. I guess my
point here is that if you take that same vehicle 3d model and those
same shaders that make it look all pretty, and compile/run them on the
360, it will look the same and run at a higher framerate than on the
PS3. This is because the PS3's gpu isn't as good as the 360's."
"I'm assuming you also know that the PS3's vertex processing units are
terrible, since each extra vertex shader input ads one cycle of delay.
Likewise, you probably also know that the only way currently around
this limitation on PS3 is to use the spu's to preprocess all geometry
by backface culling them first on cell before feeding them to the gpu.
But then you are still stuck with other PS3 gpu limitations, such as
not being able to do anti aliasing with floating point render targets
so you can't have MSAA and HDR simultaneously."
"The vertex pipeline doesn't matter when you have to waste one cycle
per vertex input. In other words, the gpu stalls untill it can fetch
all the data it needs before it can even start executing the vertex
shader. Why is this important? Because next gen games needs lots of
lookup maps to look nice, which means lots of u/v coordinates and other
data that needs to be passed to the vertex shader, ie, lots of inputs.
In RSX's case, that means gpu stalls. This is the RSX's well known
achiles heel and is well documented. The only known work around at the
moment is to use the cell spu's as another 'gpu', in this case a
culling gpu, to limit the number of verts actually sent to the RSX.
Whether or not RSX+SPU culling will help PS3 meet or exceed xenon's gpu
performance isn't known at this point."
"You don't think no msaa with floating point buffers isn't a huge
limitation? There are other color spaces, but they are useless to me. I
don't need 8/8/8/8 int. I need 8/8/8/8 float. In RSX's case you have no
choice but to use FP16 (16/16/16/16), compared to FP8 on xenon. So you
are forced to move around twice the memory on RSX if you want a
floating point buffer, which means less framerate. Just try it! I'll
assume your a game dev. Switch your PS3 game from FP16 to an 8/8/8/8
int format and see your framerate jump. Of course, you'll have to forgo
HDR on your shipping title, but you can then do msaa. Or, go back to
FP16 since HDR looks so cool, but oh ya, you then have to turn off
msaa. I just don't have these headaches on 360, but I have to deal with
them PS3."
____________________________________________________________
sounds like he knows what he is talking about and is not bullshitting.
everything he is saying fits well with what other developers have been
saying for many months, even a year or more.
