DX10 Interview w/ ATI

[jAkUp]

http://www.extremetech.com/article2/0,1697,1985159,00.asp

ET: We know that DX10 has unified pixel, vertex, and geometry shaders on the API level, but it doesn't dictate the hardware implementation—it's possible to have DX10-compliant graphics cards with discrete pixel and vertex/geometry shaders. Do you think the time is right to bring them together, or is there still some benefit from having separate units on a hardware level?

Drebin: One thing we can see even with today's games is that the amount of vertex and pixel processing even within a single frame can vary widely. Something like a sky or an object that's large and close to the camera may maximize the pixel shaders, while in the same frame, something like a character that's off in the distance could have a high polygon count but not take up too many pixels, and that may maximize the vertex shaders. Developers today have to account for all of this.

With DirectX 10, developers will have to also contend with the addition of the geometry shader, making it even harder to balance the workload. With a unified shader architecture, developers are free to make a scene how they want, focusing on pixel processing, vertex processing or geometry processing as they see fit without the restrictions of the ratio chosen for them in a discrete shader implementation. No longer are developers pixel constrained or vertex constrained; with a unified shader architecture, they become compute constrained, able to play with the total performance available to them in the GPU. Today developers are already seeing the benefits of unified shading on the Xbox360, and we're confident that those developers who have enjoyed the freedom of a unified architecture will find it hard to go back.

[Redeemed]

Okay, I'd like to hear nVidia's standpoint. Last I heard G80 wasn't going to be unified. As such, is this going to be another NV30 to some extent?

I do doubt that the G80 will be comparable to the NV30 fiasco, in fact I half expect that it will shock the sharts out of everybody. Then again, maybe nVidia predicted wrong and unified would be the way to go- hence the G80 not performing as well as the R600 in the long run.

Possible?

[tieros]

I just don't get how ATi can sell the concept of generalized hardware as a "feature".

The need for task-optimized hardware is the reason we have GPUs in the first place.

Having one large pool of generic shaders vs. two smaller pools of optimized shaders is like having a silverware drawer full of sporks :)

I hereby dub the R600 "The Spork" :D

[Redeemed]

If utilized properly unified shaders could be incredibly faster (more efficient) than what we have seen with today's hardware. But I doubt the R600 will be a very good example of this.

I have nothing wrong with either company going with unified shaders, but I just wonder why at this point. I guess ATI is probably just testing the waters, so to speak.

[jAkUp]

I just don't get how ATi can sell the concept of generalized hardware as a "feature".

The need for task-optimized hardware is the reason we have GPUs in the first place.

Having one large pool of generic shaders vs. two smaller pools of optimized shaders is like having a silverware drawer full of sporks :)

I hereby dub the R600 "The Spork" :D

Haha interesting analagy... however, the DirecX team is trying to make the GPU more generalized... kinda like CPU's.

[Lfctony]

I see a lot of potential and I believe USA is the way to go. Whether ATI get it right with the R600 is another issue. Sure, the x360 GPU was their first attempt, but we can't really compare that to current GPUs. It remains to be seen. :)

Tony.

[retsam]

I hereby dub the R600 "The Spork" :D
hahah that was great !!>...gave me a chuckle.:D

[Ristogod]

I like the sport analogy. You make a valid point.

However, the GPU may become more generalized with a USA, but it's still more specialized than a CPU is.

I guess ATI feels that the weight of emphasis should be shifted more towards generalization to find a better balance resulting in better performance. I guess it remains to be seen whether or not it will follow through.

Anyone know if this is all done through hardware on the fly though? Or does the software make the request for how they want their shaders set up for rendering a scene?

[{WOR}lock]

It's fair enough to accept that a singular unified shader (or generalized hardware as you like to put it here) could quite likely be slower than a singluar conventional shader. But the point is, that unified shaders will be balancing between what shading is needed at any given frame.

So it's not really fair to say this way is going to be faster than that way until you actually get your hands on the cards and benchmark them against each other. Until then it's all just speculation because there's no unified format
to compare with at this moment in time.

And yes, this is my first post here. But I am trying to be objectional, so don't mock me for that. ;)

[tieros]

It's fair enough to accept that a singular unified shader (or generalized hardware as you like to put it here) could quite likely be slower than a singluar conventional shader. But the point is, that unified shaders will be balancing between what shading is needed at any given frame.

So it's not really fair to say this way is going to be faster than that way until you actually get your hands on the cards and benchmark them against each other. Until then it's all just speculation because there's no unified format
to compare with at this moment in time.

And yes, this is my first post here. But I am trying to be objectional, so don't mock me for that. ;)

Welcome to the forum!

Yes, this is all very speculative, and I imagine the debate will continue even after the next generation of GPUs are released :D

My belief is that the work done to move to a unified architecture is wasted effort. My hypothesis is that in 3 years time or less, GPUs will be comprised of two or three major subsections, probably on different dies, with completely non-unified shaders.

Instead of detailing it here, I'll post something in the rumors (and wild ass guesses :D) forum sometime this week.