ATI finalised R520 clocks...

[shadow001]

Yup.

Although I like what I read about NVIDIA's unified architecture more than ATI's. Will be very interesting to see which comes out ahead.

ATI needs to hurry up and release some R520 cards! :POKE:


They don't have a unified architecture at all...Seems like Nvidia wants to keep vertex and shading engines seperate and at most unify things at the software lvl only...

[NoWayDude]

Actually,that richard huddy interview released a few days ago,did mention that the Xenos chip have 3 specific tecnical characteristics,one of which is feasable on PC graphics cards right now,the remaining 2 are possibilities for future PC cards...

Hence my question, if so, and aprentely the yelds are good, why not release Xenos for PC?
Or is it because not is all well and possibily the conversion is not that simple?
Or is it MS that does not allow this?
That is a point that i would like to read about it.Any links where they can especify more about the similarities/diferences from both architectures?

[Raje]

They don't have a unified architecture at all...Seems like Nvidia wants to keep vertex and shading engines seperate and at most unify things at the software lvl only...

"At all" right now as far as NVIDIA has let us know.

Yeah, Kirk was talking about at first only having the drivers take care of the unification. He was also talking about Homogeneous vs. Heterogeneous units in much the same way that AMD was talking about for future multicore CPUs.

The question is:
Do you have a lot of general purpose units that because of how games are programmed (Geometry load vs. Pixel Shader Load) don't ever need to be totally general purpose, or do you optimize a certain % of your units for Geometry Processing while letting them work on some Pixel Shading tasks if need be and optimize the other % for Pixel Shading while letting them work on some Geometry Processing if need be. Because transistors aren't free it might be a better idea to not go totally unified until 65nm or 45nm or when some future process technology makes it 'the way to go'.

Here's how I am looking at it and I think, from what I have read, NVIDIA is looking at it:

~Descreet VS and PS/Heterogeneous (Now) [Far Left End of the Meter]

~Unified/General Purpose Units/Homogenous (Future) [Far Right End of the Meter]

~At some point between them there is at least one stage where you aren't totally descreet and you aren't totally unified. [Some Point(s) in the Middle of the Meter]

If transistors/die size and power consumption didn't matter, I think moving directly to Unified would be the way to go. If you constrain transistors/die size and want, relatively speaking, lower power comsumption, Unified might be further off.

You can be sure that if ATI has a Unified Architecture and NVIDIA doesn't all you will hear is "ATI has unified architecture and NVIDIA is living in the dark ages with its descreet logic!" I'm all for new architectures that are "better", but there's a lot more to it than "Hey, this looks really good on paper. Lets just do it".

[Raje]

EDIT: Better source about Unified Architectures

Link (http://www.bit-tech.net/bits/2005/07/11/nvidia_rsx_interview/4.html)

"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.

"It's far harder to design a unified processor - it has to do, by design, twice as much. Another word for 'unified' is 'shared', and another word for 'shared' is 'competing'. It's a challenge to create a chip that does load balancing and performance prediction. It's extremely important, especially in a console architecture, for the performance to be predicable. With all that balancing, it's difficult to make the performance predictable. I've even heard that some developers dislike the unified pipe, and will be handling vertex pipeline calculations on the Xbox 360's triple-core CPU."

And ATI's view:

Link (http://www.bit-tech.net/bits/2005/06/10/richard_huddy_ati/3.html)

“I’d love to say yes… I’d love to say that Nvidia are going to be stuck when it comes to Longhorn. But actually I do think they will have a unified shader architecture by the time WGF2 comes around. This time around, they don’t have the architecture and we do, so they have to knock it and say it isn’t worthwhile. But in the future, they’ll market themselves out of this corner, claiming that they’ve cracked how to do it best. But RSX isn’t unified, and this is why I think PS3 will almost certainly be slower and less powerful.

“Talking to the guys at Microsoft, it’s impossible to escape the conclusion that the future is for unified pipelines, there’s no doubt.”

:fanboy: (duel)

END EDIT

http://www.extremetech.com/article2/0,1697,1745060,00.asp

Question: Are GPU architectures and Direct3D evolving toward a design where the distinction between vertex and pixel shaders essentially goes away?—davesalvator

David Kirk: For hardware architecture, I think that's an implementation detail, not a feature.

For sure, the distinction between the programming models and instruction sets of vertex shaders and pixel shaders should go away. It would be soooo nice for developers to be able to program to a single instruction set for both.

As to whether the architectures for vertex and pixel processors should be the same, it's a good question, and time will tell the answer. It's not clear to me that an architecture for a good, efficient, and fast vertex shader is the same as the architecture for a good and fast pixel shader. A pixel shader would need far, far more texture math performance and read bandwidth than an optimized vertex shader. So, if you used that pixel shader to do vertex shading, most of the hardware would be idle, most of the time. Which is better—a lean and mean optimized vertex shader and a lean and mean optimized pixel shader or two less-efficient hybrid shaders? There is an old saying: "Jack of all trades, master of none."

Also, not that it really belongs here, but I found this while trying to track down links to where Dr. Kirk talks about Unified Architecture stuff:

http://www.neoseeker.com/Articles/Hardware/Features/geforce6800/10.html

David Kirk: As we go forward with GPU evolution, we will continue to make GPUs more powerful, more programmable, and more flexible. GPUs are already the most "bang for the buck" in terms of floating point computation per dollar, but we can do a lot better. We want to make GPUs useful for producing game geometry, and calculating game physics and AI/behavior as well.

[shadow001]

EDIT: Better source about Unified Architectures

Link (http://www.bit-tech.net/bits/2005/07/11/nvidia_rsx_interview/4.html)



And ATI's view:

Link (http://www.bit-tech.net/bits/2005/06/10/richard_huddy_ati/3.html)



:fanboy: (duel)

END EDIT

http://www.extremetech.com/article2/0,1697,1745060,00.asp



Also, not that it really belongs here, but I found this while trying to track down links to where Dr. Kirk talks about Unified Architecture stuff:

http://www.neoseeker.com/Articles/Hardware/Features/geforce6800/10.html



GPU's with Physics and AI?....Intel and AMD are going to have a fit if their dual core CPU's,which will soon(~2007),going to become triple and even quad core cpu's,aren't going to get used in the one thing that a large part of people still upgrade their computers for...Gaming,since PC's as they are right now,are fast enough for everything else for most people....


Seems to me that it's getting to the point where there's far too much redundancy in PC's,especially those with multiple CPU cores,wich will become the norm rather than the exception sooner or later,and video card makers trying to add physics and AI on their GPU's,while yet other companies(ageia) wants to sell dedicated physics boards,and i've heard of yet another company who wants to also make dedicated AI chips for gaming(can't remember the name)...

[Raje]

GPU's with Physics and AI?....Intel and AMD are going to have a fit if their dual core CPU's,which will soon(~2007),going to become triple and even quad core cpu's,aren't going to get used in the one thing that a large part of people still upgrade their computers for...Gaming,since PC's as they are right now,are fast enough for everything else for most people....


Seems to me that it's getting to the point where there's far too much redundancy in PC's,especially those with multiple CPU cores,wich will become the norm rather than the exception sooner or later,and video card makers trying to add physics and AI on their GPU's,while yet other companies(ageia) wants to sell dedicated physics boards,and i've heard of yet another company who wants to also make dedicated AI chips for gaming(can't remember the name)...

There will always be a need for a fast general purpose CPU. Especially for games because the code that makes games work doesn't run well on non-speculative, special purpose, parallel or vector processors (yet at least and even when the industry reinvents itself having a fast, speculative, general purpose, multi-core CPU will be much easier to code for). You also need to be able to run the drivers to tell the GPU, PPU, and AIPU(?) what to do.

Look at Unreal Engine 3. UE3 is going to use a thread for sound, Multiple threads for Physics via the AEGIA SDK (was Novadex and now PhysX(?) I think), a thread for gameplay, a thread for graphics calls. On top of that NVIDIA is supposed to have Multi-threaded drivers coming (supposedly in the 80s driver releases). Also consider all the threads that the OS has running. There's a lot that a CPU will be doing and can be doing.

The way I think about it is the CPU, GPU, PPU, AIPU(?) all work together and the CPU is what makes them work together. Having a PPU and an AIPU(?) will make the load on the CPU less (just like having a dedicated hardware rendering device (GPU) took the burden of rendering off the CPU) and then game designers will utilize that extra CPU power for something else.

As for descreet GPU, PPU, and AIPU, I'd rather have everything on one Programmable GPU. Seems a GPU that deals with all the geometry anyway could easily detect collisions. Sure it might cost more if you only want a GPU -- obviously because you need more die space for the GPU, PPU, AIPU or the functional units in the GPU to make all that work -- but it would be worth it to not need to buy a PPU, a GPU and a AIPU each of which would probably need a fast PCIe slot and because having all on the same device could speed things up. Also, the cost of having all on one board should be less than having each on their own board because you use less materia and could use the same memory controller/RAM.

Sorry about the rant :o

If you want a way to utilize multiple cores for single threaded work, Intel is working on that too. ;) Turning Single into Multi-Threaded with Speculative Threading (http://www.anandtech.com/tradeshows/showdoc.aspx?i=2507&p=7)

As you can see, offering in many cases a 2 - 3x performance improvement is nothing short of impressive. But keep in mind, this project is in its very early stages of research and as promising as this looks, it may take 5 - 10 years for the research to make its way into the real world.

[nutball]

I can't help but feel that there's a parallel between the way the gaming PC is going and the boom-time of high-performance computing (60's & 70's basically). That era spawned all sorts of weird and wonderful hardware designed to accelerate very specific types of algorithm (vector processors being the most obvious).

Look at the Top 500 today, it's dominated by Linux clusters. Even high-end general-purpose processors have been forced into a niche. Vector processors still exist, but you can probably count the world population of them on the fingers of one hand (after a nasty accident with a band-saw).

The lesson is that although dedicated hardware can accelerate a specific task very effectively, it's also by its nature very inflexible (I'm not talking about programmability v. non-programmabilty, I'm talking about the next level up). It's fine so long as you're trying to do what it's designed for, otherwise you're in a world of hurt (hard to program and/or limited performance gain). Computing in a heterogeneous environment is just plain time-consuming for programmers.

Whilst the existence of standard APIs (DirectPhysics, DirectAI) might help the application programmer exploit the hardware, they could be *really* bad for innovation. If as a game developer I came up with super-whizzy new physics algorithm... but it cuts across the assumptions made by Ageia in the design of their chip so I see no benefit from it... what would I choose to do?