Nvidia’s Transition to Unified Shader Architecture Will Be Gradual

[Ubermann]

Nvidia Corp.’s chief scientist David Kirk has once again reaffirmed his stance on the unified shader processors by claiming that shortly from now there may be no point to implement them and said that the company would be “gradually” going to the unified architecture, the approach taken by arch-rival ATI Technologies.

Read more at Xbitlabs (http://www.xbitlabs.com/news/video/display/20060418170458.html)

[Cobalt]

How the hell is that going to work? Unified shaders are part of the DX10 spec. You can't go halfway on this cause MS doesn't want another FX series on its hands when it comes to API development. I mean come on, does nV even remember that far back? This stinks of cost cutting to me but it'll probably end up cutting their profits instead.

[DilTech]

Well, it's official....NVidia have lost their minds.

Time to go buy ATi stock.

[ahmad]

This stinks of cost cutting to me but it'll probably end up cutting their profits instead.

If the price is right it will sell. And they still have the mainstream market by its balls, unless ATI improves on their situation.

[Cobalt]

The problem is that the "right price" will be so low that they couldn't support themselves. ATi's postions is defo improving and if I ever go into PCWorld for any reason (shudders) then 75% of the PCs without intel will have ATi on the case.

They won't get much support from software devs if they do another cut and paste card either so we'll see a decline in products "made for nVidea" or whatever slogan they put on games. Thinking about it, I can't remember the ATi one either.

[kemist]

How the hell is that going to work? Unified shaders are part of the DX10 spec. You can't go halfway on this cause MS doesn't want another FX series on its hands when it comes to API development. I mean come on, does nV even remember that far back? This stinks of cost cutting to me but it'll probably end up cutting their profits instead.

For a while it may make sense to go this way. Remember the old phrase jack of all trades master of none? I would bet money they are just gonna make a scheduler that will accept commands for unified api and just convert to a non-unified format, sending the commands to the correct unit and possibly compiling them at the end to make it appear unified to the system. If this works and they have dedicated units it could end up being faster than unified depending on the scene being rendered (likewise unified may be quicker in some scenes than non-unified).

[turtle]

For a while it may make sense to go this way. Remember the old phrase jack of all trades master of none? I would bet money they are just gonna make a scheduler that will accept commands for unified api and just convert to a non-unified format, sending the commands to the correct unit and possibly compiling them at the end to make it appear unified to the system. If this works and they have dedicated units it could end up being faster than unified depending on the scene being rendered (likewise unified may be quicker in some scenes than non-unified).

That, is a very awesome post...Because you said what I was thinking in a much more elegant fashion than I ever could. Props for that.

Exactly. For example...What if there is 48 pixel shaders, and 16 vertex shaders...Equalling 64 shaders, on G80. That truely makes you think how that would stack up to a 64 unified shader part from ATi if either could be converted either way (dx10->dx9 or vice versa). It would come down to efficiency, and if getting software companies to create along those set variables. Considering I don't doubt the $ power of TWIMTBP, it makes you wonder if something like that holds true, if ATi will be recreating that ratio anyway even if it is with unified shaders. If that is the sweet spot, and nvidia can get it's "matured tech" running more efficient dx9/10 mix than ATi can get it's "new tech"...Heck, nvidia could have something. Yes, that is a lot of "ifs", but it does make sense...kinda sorta.

The question is, would the instances of variables outside of the set shaders for such an nvidia product (3p:1v) show a huge advantage on ATi's hardware because of it's on-the-fly vertex/pixel computation, or could nvidia just get it running faster/more efficient per clock to compensate?

It does make one wonder, that's for sure.

That being said, I still expect G80 to be 24rops/24tmus, and R600 to be 32/32. If the products end up being closer or exactly equal in that respect (Nvidia ends up having more or ATi less than I think), it could be very interesting. It's also very possible. ATi for instance, may keep it's 3:1 ratio of xenos/R580 and have 24rops and 24tmus, if it has 72 shaders, following the ratio of R580. Nvidia could very-well be at 32/48, effectively a 2xG70, which when seeing how small the die size is for G70...Seemingly very possible at 80nm. Unlikely that high of rops/tmus , but possible.

I dunno. Just complete speculation...Who knows. I still expect R600 to rule the school, but nvidia may indeed make this thing float, even if the old dx9 way of shading is a sinking ship.

[Richteralan]

This is a very good way to transition to Unified.

Remember DX10 is not necessarily mean unified in HARDWARE.

[Ubermann]

I think ATI's risk taking will some day pay off in nice boost in performance.

[STEvil]

they caught nV with their pants down with the 9x00 series.

If we see a repeat i'll laugh :D

[MaxxxRacer]

For example...What if there is 48 pixel shaders, and 16 vertex shaders...Equalling 64 shaders, on G80. That truely makes you think how that would stack up to a 64 unified shader part from ATi if either could be converted either way (dx10->dx9 or vice versa). It would come down to efficiency, and if getting software companies to create along those set variables. Considering I don't doubt the $ power of TWIMTBP, it makes you wonder if something like that holds true, if ATi will be recreating that ratio anyway even if it is with unified shaders. If that is the sweet spot, and nvidia can get it's "matured tech" running more efficient dx9/10 mix than ATi can get it's "new tech"...Heck, nvidia could have something. Yes, that is a lot of "ifs", but it does make sense...kinda sorta.

Here is the thing though. If ATI does the unified shader correctly, meaning that it can adress the pixel and vertex processing loads correctly on the fly, there is no way that a non unified shader can beat it. It would be optimizied for each frame, giving it extra effeciency and not wasting any of the shader processors as is done all the time right now.

I believe that the unified shader will allow for much more balanced performance, meaning that the card will not suck in one game and be amazing in another.. It will be closer to amazing most of the time.. atleast thats the theory.

[nn_step]

they caught nV with their pants down with the 9x00 series.

If we see a repeat i'll laugh :D
That'll only happen if and only IF Windows pulls a fast one like they did with DirectX 9

[Starscream]

Here is the thing though. If ATI does the unified shader correctly, meaning that it can adress the pixel and vertex processing loads correctly on the fly, there is no way that a non unified shader can beat it. It would be optimizied for each frame, giving it extra effeciency and not wasting any of the shader processors as is done all the time right now.

I believe that the unified shader will allow for much more balanced performance, meaning that the card will not suck in one game and be amazing in another.. It will be closer to amazing most of the time.. atleast thats the theory.

true unified is faster then anything else but will the ATI cards also be able todo this balancing for Dx9 games wich werent developed to work with unified architekture cards?

cause tbh from their stand (ATI/Nvidia) its al about whos faster in DX9 games the R600 or G80.
cause at the time those 2 cards apear they will be tested mainly in Dx9 games.

@Kemist.
there used to be a small rumor that the G80 had 2 cores.
maybe a 2nd small core todo this scheduling work.

[onewingedangel]

Prehaps everyones misreading the statement, prehaps it means that whilst the 8800 wil use unified shaders, the rest of the product line 8600,8200 etc. will remain using a fixed pipeline for the time being, hence there being a gradual move to unified shading, but not having a card be a bodge job.

[ahmad]

Here is the thing though. If ATI does the unified shader correctly, meaning that it can adress the pixel and vertex processing loads correctly on the fly, there is no way that a non unified shader can beat it. It would be optimizied for each frame, giving it extra effeciency and not wasting any of the shader processors as is done all the time right now.

I believe that the unified shader will allow for much more balanced performance, meaning that the card will not suck in one game and be amazing in another.. It will be closer to amazing most of the time.. atleast thats the theory.

Deja vu. Hardware implementation > Emulation any way you cut it. DX9 won't be affected, but DX10 will depending on how powerful the hardware is and how dependant all of the DX10 features are on this. Nvidia could go and talk to programmers: "Don't do this the dx10 way, do it our way".

Does this sound familiar to anyone at all? DX8.1 hacked to work with DX9? Well lets hope they decide to make the transition stage a short one.

Prehaps everyones misreading the statement, prehaps it means that whilst the 8800 wil use unified shaders, the rest of the product line 8600,8200 etc. will remain using a fixed pipeline for the time being, hence there being a gradual move to unified shading, but not having a card be a bodge job.

Yes everyone pretty much understood this. First G80 cards will all be using a fixed number of pipes and eventually nvidia will transition to a proper unified shader design. No one is arguing that. The fact that they are deciding to do this transition stage during the release of DX10 isn't something that will... how shall we say this? blow the market away. And besides, who knows how long it will take them to fully make the transition?

[Starscream]

Prehaps everyones misreading the statement, prehaps it means that whilst the 8800 wil use unified shaders, the rest of the product line 8600,8200 etc. will remain using a fixed pipeline for the time being, hence there being a gradual move to unified shading, but not having a card be a bodge job.

could be true, i remember there being a post a week orso ago stating that the G8x budget and mid-range cards wouldnt be unified due to them cards else not being powerfull enough in Dx9 aplications vs true dx9 cards.


if non of the G8x cards will be fully unified wont hurt Nvidia to much tbh.
Maybe their marketing will b hurt by it but the rest i doubt it.

As it will take a long time till the consumers have moved to Vista.
It will also take a while till a good bunch of Dx10 games are around
Also someone said that Nvidia onboard GPU is able todo Areo (or howver its written) mode of Vista (wich is like gonna be a good seller).

so IF non of the G8x cards r unified theyl have a hard time with marketing it but thats about it i guess for the first year.

[DilTech]

Prehaps everyones misreading the statement, prehaps it means that whilst the 8800 wil use unified shaders, the rest of the product line 8600,8200 etc. will remain using a fixed pipeline for the time being, hence there being a gradual move to unified shading, but not having a card be a bodge job.


Actually, this makes perfect sense, doubtful after this new, but makes perfect sense.

My question is, if they're going "hybrid", what exactly do they mean? If it's all fixed function, that'd make it nothing close to a hybrid. A Hybrid card would have to be a bit of both, right?

My theory?
The standard pixel shaders(48 in this case) are fixed function.
The Vertex/Geometry Shaders are unified, meaning they can go Pixel, Vertex, Geometry on the fly.

This would mean, technically, that it can do anything the ati card can do since no game will need more than 16 vertex or geometry shaders at any given time, while no game currently would use less than the 48 pixel shaders. If they use more, the 16 unified now switch over to pixel shaders.

If that is correct, this might not be so bad after all.... It'd be a beast in DX9 and DX10 without giving up anything to a full unified card(except in VERY extreme situations). It's going to be very rare than a game needs more vertex rendering power than pixel shader power.

[Cobalt]

I'm still doubtful. Even if the transition is quick they would need a major overhaul of the architecture to work with unified shaders even if they lay the ground work. They will either have to give it all the necessary bits to do unified but then not use them (waste of die space), or actually have two different architectures even if they are compatible (waste of money cause they could have just concentrated more resourses on getting the DX10 parts to perform on par with ATi).

[Starscream]

if this G80 is a hybrid i wouldnt quickly say their next chip would require a major overhaul.
Nvidia like ATI has been working on a unified architekture for years i doubt that theyve waisted all that time on a hybrid design wich can only last 1 gen.

Cause pumping alot of money researching new tech for a hybrid design wich has a short life is stupid (wich can only last 1 gen) and also alot fo cash and time in unified (wich will last alot longer)

if G80 is hybrid i think its a basic (clever) mix of tech they already had or they got while working on unified.

so to be short if its a hybird im shure that they have thought of all the things needed to make it work (Dx10 and games) as they know thats were it has to work good.

[Aynjell]

Well, it's official....NVidia have lost their minds.

Time to go buy ATi stock.

I disagree, bud. nVidia has something much cooler up thier sleeves, and that is what they are hinting at. They are just letting the world know that this is an ugly hack that is implementable now (and must be met because well... everyone that matters is clamouring over unified shaders because MS demands it) and will be outmoded shortly after by something possibly more complex, but ultimately more appropriate.

Intel seems to think similar thoughts about SOI, and yet they remain competitive (or is conroe gonna be SOI?)...

[[XC] leviathan18]

or maybe we are reading the wrong way g80 hybrid card for this summer to create a base of ppl with vista semi ready GPU and a proper card in january/vista release.... that makes sense too

[Flak Monkey]

Actually, this makes perfect sense, doubtful after this new, but makes perfect sense.

My question is, if they're going "hybrid", what exactly do they mean? If it's all fixed function, that'd make it nothing close to a hybrid. A Hybrid card would have to be a bit of both, right?

My theory?
The standard pixel shaders(48 in this case) are fixed function.
The Vertex/Geometry Shaders are unified, meaning they can go Pixel, Vertex, Geometry on the fly.

This would mean, technically, that it can do anything the ati card can do since no game will need more than 16 vertex or geometry shaders at any given time, while no game currently would use less than the 48 pixel shaders. If they use more, the 16 unified now switch over to pixel shaders.

If that is correct, this might not be so bad after all.... It'd be a beast in DX9 and DX10 without giving up anything to a full unified card(except in VERY extreme situations). It's going to be very rare than a game needs more vertex rendering power than pixel shader power. I really like that idea and it actually seems pretty likely that they would use an approach like this. The only thing I worry about is the number of transistors that would be needed to facilitate something like this.

[kemist]

Here is the thing though. If ATI does the unified shader correctly, meaning that it can adress the pixel and vertex processing loads correctly on the fly, there is no way that a non unified shader can beat it. It would be optimizied for each frame, giving it extra effeciency and not wasting any of the shader processors as is done all the time right now.

I believe that the unified shader will allow for much more balanced performance, meaning that the card will not suck in one game and be amazing in another.. It will be closer to amazing most of the time.. atleast thats the theory.

This is why i said it could depend and either way could be faster. In an extreme situation favoring all PS ops or all VS ops for example unified should theoretically be better. Under most circumstances there isnt going to be an extreme like that and you will have a mixture of both. If you set up your units so that they are in the sweet spot for whatever ops generally come in you will be just as fast as unified under most circumstances. However, it is possible that a dedicated unit is faster than a unified unit at its given op, but even if it is not it probably requires a whole lot less silicon to implement, your main cost would be in a scheduler that could handle sorting out the ops and putting it all back together but unified requires a complicated scheduler anyway. So if a non unified approach requires less silicon you can potentially implement more units or have a smaller die.

I think its going to be kinda like the situation now: R580 is big and complicated but smart and able to compete with only 16 pipes (or insert R520 if you wanna argue about the extra ps units). Whereas g70 is smaller but has brute force to be competitive.

p.s. that probably isnt the clearest post so if any of that needs clarified let me know

[onewingedangel]

If the scheduler cant change things on the fly with any degree of efficiency, theres still the benefit of adjusting the ratios to best suit the game, if not the scene exactly. So if a game prefers a heavier pixel shader ratio have it boot at 16/48 or if its less pixel shader dependant have it boot at 24/40 etc.

[DilTech]

I really like that idea and it actually seems pretty likely that they would use an approach like this. The only thing I worry about is the number of transistors that would be needed to facilitate something like this.

Less than it'd take for all the shaders to be unified, that's for sure. Alot less.

This approach makes sense, as NVidia has patented unified shaders(as I showed in my first post in this thread) so it's obvious they're going to use it for something.

I could be wrong about this, again, it's just my guess. It's just what makes the most sense. Why else would they call it a "hybrid" if it wasn't part fixed function and part unified? It'd also make it a "gradual" change, now wouldn't it?

[Starscream]

if it would take alot less transistors could it be that theyl try to price it alot lower then the R600?

[Thorry]

Well I get it but it sounds like B/S...

How can you gradually go into a unified shader architecture. Either you have unified shaders or you don't, there is nothing in between...

[ahmad]

This is why i said it could depend and either way could be faster. In an extreme situation favoring all PS ops or all VS ops for example unified should theoretically be better. Under most circumstances there isnt going to be an extreme like that and you will have a mixture of both. If you set up your units so that they are in the sweet spot for whatever ops generally come in you will be just as fast as unified under most circumstances. However, it is possible that a dedicated unit is faster than a unified unit at its given op, but even if it is not it probably requires a whole lot less silicon to implement, your main cost would be in a scheduler that could handle sorting out the ops and putting it all back together but unified requires a complicated scheduler anyway. So if a non unified approach requires less silicon you can potentially implement more units or have a smaller die.

I think its going to be kinda like the situation now: R580 is big and complicated but smart and able to compete with only 16 pipes (or insert R520 if you wanna argue about the extra ps units). Whereas g70 is smaller but has brute force to be competitive.

ATI could make a hack for the r580 to support DX10, because from the article its clear nvidia is making something very similar to the x1k design (minus the mem controller probably). Heck what are the chances they haven't already done that?

Also you talk about "extreme" situations when you don't realize these cards are going to be compared head to head, and everyone is going to see a difference in numbers. Even if its miniscule. A 1 FPS advantage for either side will result in them being crowned the king.

Also there is going to be more pressure on nvidia because ATI had the "best IQ" title, so everyone is going to be looking forward to how nvidia will take that or match it.

What I am looking forward to is some serious improvements from both sides. I want to see OpenGL that matches competition from ATI, and I want nvidia to do what they say they will...

[POGE]

I just hate the fact that the cutting edge dx9 cards are going to suck at dx10 games. :(