r520 has unified shader architecture?

[kilmas]

the register say that the r520 has unified shader architecture,can that be true?

i have read that we must wait for r600 to come out next fall before we get to see unified shader architecture on pc.

http://www.theregister.co.uk/2005/09/23/ati_r520_launch_date/

[Raje]

I think the problem is that most people can't deal with or comprehend ATI having a VPU in Xbox360 and one for the PC world that use different technology.

I seriously doubt that R520 is unified. If anything, R520 has some new technology relative to R420, but not something one would call a "Unified Architecture" on the hardware level.

[Rakeesh]

I am not a 3D hardware guru...what is implied by "unified shader architecture?" In other words, how does having that feature set those cards apart from cards that don't have this?

[Sazar]

I am not a 3D hardware guru...what is implied by "unified shader architecture?" In other words, how does having that feature set those cards apart from cards that don't have this?

Instead of having different sets of shaders for different things (ie vertex) one set allows you to work both sets.

Added benefit is the efficiency of utilization of those units.

I think I said it right :D

[Subtestube]

It should be noted that although utilisation efficiency goes up, performance per unit should (theoretically) go down. Basically, more special purpose typically = better perf. And vice versa.

[jolle]

the point is mainly load balancing and removing bottlenecks.. like you wont have any PS units standing idle if there is a bigger Vertex Shader load..
makes more sense on consoles atm, since its a fixed hardware platform, and the devs gets a little bit more freedom with it.

I dont think the R520 is acctually using unified shaders yet..
Its optional even for WGF2.0 I think..

[Subtestube]

Its optional even for WGF2.0 I think..

My understanding is that it's compulsary at an API level - i.e., Vertex and Fragment shaders must have equivalent feature sets. At the hardware level, to my knowledge, it's not required. It's possible the R520 could have 'unified' shaders in feature set and not at the hardware level - once you hit shader model 3.0, you're getting pretty close to that in any case.

[Zelda_fan]

I think the problem is that most people can't deal with or comprehend ATI having a VPU in Xbox360 and one for the PC world that use different technology.

Why would ATI create an advanced unified shader GPU for XBOX 360 then use outdated technology for a PC addin card? That makes no sense.

[Razor1]

Why would ATI create an advanced unified shader GPU for XBOX 360 then use outdated technology for a PC addin card? That makes no sense.


Because designing a Unified Architecture is a bit more costly then going the tried and tested route, also testing it out in the xbox 360 is perfect since consoles gpu's don't have any other competition like in desktops, just incase this new architecture doesn't have a good scheduler that will give it better efficiency then today's desktop gpu's.

Most likely the r520 isn't a unified archecture but something in the middle of what we have today and unified shaders. They will most likely have pixel shaders and vertex shaders decoupled from each other, so data doesn't need to flow from vertex shader to pixel shader like todays GPU, giving it more efficiency during certian shader operations such as texture look ups that are created via a buffer.

[Dazz]

I think the unified shader architecture is the way to go NV have said they they hated it but are ending up going the route also. So rather the have a 24pipe shader card with 8 Vertex you can have a 32pipe card that can dynamicly change. SO if a game is pixel shader dependent it can use say 30 pipes for shaders and 2 for Vertex. If the Game is Vertex depended and not shader then it can use 24 pipes for Vertexes and 8 for pixel shading. It does add more flexablity for developers.

[Raje]

I think the unified shader architecture is the way to go NV have said they they hated it but are ending up going the route also.

Links?

If this information is available somewhere, I would really like to read about it.

I haven't read anything saying that NV has planned on going unified on the hardware level yet other than something along the lines of it might be the way to go in the future and When unified will be the better option on the hardware level, then we will do it.

The feeling I have from everything I have read is that NVIDIA is going to wait on a hardware unified architecture in favor of an architecture that makes better use of die size (transistors), functional units, power, and has lower heat output.

I'm certain that NVIDIA will support a unified API on at least the Driver level.

So rather the have a 24pipe shader card with 8 Vertex you can have a 32pipe card that can dynamicly change. SO if a game is pixel shader dependent it can use say 30 pipes for shaders and 2 for Vertex. If the Game is Vertex depended and not shader then it can use 24 pipes for Vertexes and 8 for pixel shading. It does add more flexablity for developers.

That 32 pipe GPU that can dynamically change will be, overall, less efficient than a descreet 8 vertex shader, 24 pixel shader GPU though (assuming your graphical load calls for 8 vertex shaders and 24 pixel shaders). The benefit from the unified architecture on the hardware level is that you can have, in theory (with a 32 unit unified GPU), geometry shading power between 0 and 32 and pixel shading power between 0 and 32. Where Geometry Power = Total Number of Units - Active Pixel Shader Units and Pixel Shading Power = Total Numer of Units - Active Geometry Shading Units.

But how often will pixel shading power reach 0 and how often will geometry shading power reach 0? How often will they both reach 2, 4, or N? If Geometry or Pixel power is never less than N or even less than N only a small fraction of the time, then you could have N descreet, optimized VS or PS and the rest of the units could be general purpose. A GPU set up this way would be more efficient because the optimized units would probably clock higher, take up less die space, and be better at their jobs than the hybrid/general purpose units.

What worries me is that now people are busy coming up with theoretical geometry and pixel shading loads in order to show how useful a unified architecture is eventhough those theoretical loads aren't indicative of actual games. They are so set on "Unified is newer and therefore must be better" that they are over looking the true problem:
To quote Dr. Kirk:
Debating unified against separate shader architecture is not really the important question. The strategy is simply to make the vertex and pixel pipelines go fast. The tactic is how you build an architecture to execute that strategy. We're just trying to work out what is the most efficient way.

[GlowStick]

Unified shader right now is just a buzzterm that anyone can use. Basicly you can use it as a blanket term for anything. Its the 'Per Pixel Lighting' of game engines, but for video cards. They are all per pixel lighting :o

[Raje]

Unified shader right now is just a buzzterm that anyone can use. Basicly you can use it as a blanket term for anything. Its the 'Per Pixel Lighting' of game engines, but for video cards. They are all per pixel lighting :o

:confused:

If your hardware is unified there is a physical difference between it and non-unified hardware. Better or not the architecture is different.

EDIT: Then again I guess if you support a unified API via drivers your Graphics Solution could be called "unified". ;) Maybe that's what you meant.