Support NeoGAF

[gofreak]

http://www.bit-tech.net/bits/2005/07/11/nvidia_rsx_interview/1.html

Not very enlightening, unfortunately. All the interviewer seemed to care about re. RSX is if it
was going to be faster than the 7800GTX, which we already knew the answer too.

The comments on unified shaders are interesting though (particularly the bolded, if true).

"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."

 

[Wunderchu]

a little bit of discussion @ Beyond3D: http://www.beyond3d.com/forum/viewtopic.php?t=24800

(currently not very much , but I have a hunch it will pick up during the day)

 

[mrklaw]

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."

If that was supposed to be his swipe at ATI, then its a stupid one. Oh, tee hee, unified shaders are so shit, that developers will do all the vertex stuff on the CPU, leaving THE ENTIRE GPU TO DO PIXEL SHADERS. Heh, anyone can do that.....well.....except us.....


Doh!

 

[Wunderchu]

If that was supposed to be his swipe at ATI, then its a stupid one. Oh, tee hee, unified shaders are so shit, that developers will do all the vertex stuff on the CPU, leaving THE ENTIRE GPU TO DO PIXEL SHADERS. Heh, anyone can do that.....well.....except us.....


Doh!

hehe :lol

 

[Nostromo]

The most interesting thing that Mr Kirk says in that interview is, imho, that RSX doesn't have any support for MSAA with FP render targets, as G70.
Supersampling AA here we go

 

[gofreak]

If that was supposed to be his swipe at ATI, then its a stupid one. Oh, tee hee, unified shaders are so shit, that developers will do all the vertex stuff on the CPU, leaving THE ENTIRE GPU TO DO PIXEL SHADERS. Heh, anyone can do that.....well.....except us.....


Doh!

The problem is if you're being forced to do this in order to maintain performance.

It would presumably be Kirk's point that in such instances, your unified shaders are a waste because you're just doing one task (pixel shading) on an architecture designed for two - and you'd be better off with a solo pixel-shader or in the bigger picture with a more traditional dedicated architecture.

Think about it this way - with unified shaders, your performance per shader vs a dedicated shader is not going to be as high. How much of a hit you take is unknown, but we can work under the fairly safe assumption that there is some hit. With a dedicated architecture, usually maybe 75% of your chip is made up of dedicated pixel shaders. So assuming all else is equal, you're unified shaders all working on pixel shading, individually can offer no less than 75% of the performance of a dedicated shader, or else you'll under perform vs the dedicated architecture. And you're still losing the vertex shading capability and eating CPU cycles to compensate (and on X360 I think the CPU hit would be significant).

I don't know if what Kirk is saying is true, however, about some devs doing no vertex shading on the GPU. Funnily enough, though, there had been speculative discussion some time ago on B3D about how unified shaders may not be so good for vertex shading (expected loss of per-shader efficiency aside).

 

[mrklaw]

I know, I was just being an arse


Plus I didn't feel like pointing out that if previous specs are to be believed, even if Xenos is doing only pixel shading, RSX still beats it *and* has vertex shaders ready for action.

 

[midnightguy]

the long-term future of the Sony-Nvidia partnership looks strong - PSP2 and PS4 far off beyond the horizon.

 

[Chiggs]

I don't know if what Kirk is saying is true, however, about some devs doing no vertex shading on the GPU. Funnily enough, though, there had been speculative discussion some time ago on B3D about how unified shaders may not be so good for vertex shading (expected loss of per-shader efficiency aside).

I've seen the exact opposite argued, as well. A shame we really won't know much until development seriously ramps up on both machines.

 

[llTll]

i am surprised Sony/Nvidia didnt talk much about RSX so far.

i want to compare it to ATI`s one

 

[dorio]

The problem is if you're being forced to do this in order to maintain performance.

It would presumably be Kirk's point that in such instances, your unified shaders are a waste because you're just doing one task (pixel shading) on an architecture designed for two - and you'd be better off with a solo pixel-shader or in the bigger picture with a more traditional dedicated architecture.

Think about it this way - with unified shaders, your performance per shader vs a dedicated shader is not going to be as high. How much of a hit you take is unknown, but we can work under the fairly safe assumption that there is some hit. With a dedicated architecture, usually maybe 75% of your chip is made up of dedicated pixel shaders. So assuming all else is equal, you're unified shaders all working on pixel shading, individually can offer no less than 75% of the performance of a dedicated shader, or else you'll under perform vs the dedicated architecture. And you're still losing the vertex shading capability and eating CPU cycles to compensate (and on X360 I think the CPU hit would be significant).

I don't know if what Kirk is saying is true, however, about some devs doing no vertex shading on the GPU. Funnily enough, though, there had been speculative discussion some time ago on B3D about how unified shaders may not be so good for vertex shading (expected loss of per-shader efficiency aside).

That's strange since one of the big touted features of the ps3 was to have the cell processor do the vertex processing now Kirk is saying that's a bad thing. What I take from the interview is that it is quite impressive what ATI managed to do with the xgpu. They managed to include 2 big features in which even Kirk admits is in the future, unified architecture and the ability to do AA along with HDR rendering on a consumer level platform. MS owes those guys big time.

 

[gofreak]

That's strange since one of the big touted features of the ps3 was to have the cell processor do the vertex processing now Kirk is saying that's a bad thing.

You're right, it had been speculated that PS3 would put all vertex shading on the CPU and dedicate the GPU to pixel shading (however this seems highly unlikely now). There are a number of big differences however between this and what the described scenario on X360:

1) It was speculated that the PS3 GPU would be using dedicated pixel shaders, not unified shaders.

2) The reason for going with unified shaders is to auto-balance between vertex and pixel work and gain greater utilisation - if you're only doing one of those tasks (pixel shading) you lose your biggest advantage and gain none of the efficiency of dedicated hardware

3) There's much more power on Cell for this kind of work than on Xenon - you could do vertex processing on Cell and still have a lot to spare for other tasks. The same can't be said of Xenon, unfortunately.

While that kind of setup, with a dedicated pixel processor, might make sense in PS3, I don't think it'd fit very well at all in X360. You'd be giving up Xenos's strong points, and taking too many hits elsewhere in the process.

Does that mean no games might do that? No. I mean if Kirk is to be believed some are doing that. But I can't fathom what would be so up with Xenos as to make that a better alternative.

 

[dorio]

You're right, it had been speculated that PS3 would put all vertex shading on the CPU and dedicate the GPU to pixel shading (however this seems highly unlikely now). There are a number of big differences however between this and what the described scenario on X360:

1) It was speculated that the PS3 GPU would be using dedicated pixel shaders, not unified shaders.

2) The reason for going with unified shaders is to auto-balance between vertex and pixel work and gain greater utilisation - if you're only doing one of those tasks (pixel shading) you lose your biggest advantage and gain none of the efficiency of dedicated hardware

3) There's much more power on Cell for this kind of work than on Xenon - you could do vertex processing on Cell and still have a lot to spare for other tasks. The same can't be said of Xenon, unfortunately.

While that kind of setup, with a dedicated pixel processor, might make sense in PS3, I don't think it'd fit very well at all in X360. You'd be giving up Xenos's strong points, and taking too many hits elsewhere in the process.

Does that mean no games might do that? No. I mean if Kirk is to be believed some are doing that. But I can't fathom what would be so up with Xenos as to make that a better alternative.

Of course alot depends on the cpu requirements for a game and the performance of the xcpu at vertex operations. In terms of doing all your pixel shading on the gpu though, you obviously wouldn't have the same number of units do this on unless you had the unified architecture because those 8 vertex units would be wasted because they would have nothing to do.

 

[gofreak]

In terms of doing all your pixel shading on the gpu though, you obviously wouldn't have the same number of units do this on unless you had the unified architecture because those 8 vertex units would be wasted because they would have nothing to do.

You could spare your CPU using those 8 vertex shaders

As for pixel shading, since you seem to be comparing a unified architecture to a dedicated architecture with 8:24 ratio of vertex to pixel shaders, as I said it's very questionable if your greater number of execution units will overcome lost performance within the units due to their less dedicated nature. How much is lost relative to a dedicated unit remains unknown, of course, but there is some X factor there to overcome. Although it may be a simplification you'd be comparing 75% of a very dedicated chip to 100% of a more general chip. How that would play out is anyone's guess right now, but I wouldn't place money on a significantly favourable outcome (especially if you introduce further differences...like say, a clockspeed differential ).

Even if you managed to eke out a small performance win on pixel shading vs the dedicated unit, you're still losing so much elsewhere, and it's questionable if anyone would notice your gains.

 

[dorio]

You could spare your CPU using those 8 vertex shaders

As for pixel shading, since you seem to be comparing a unified architecture to a dedicated architecture with 8:24 ratio of vertex to pixel shaders, as I said it's very questionable if your greater number of execution units will overcome lost performance within the units due to their less dedicated nature. How much is lost relative to a dedicated unit remains unknown, of course, but there is some X factor there to overcome. Although it may be a simplification you'd be comparing 75% of a very dedicated chip to 100% of a more general chip. How that would play out is anyone's guess right now, but I wouldn't place money on a significantly favourable outcome (especially if you introduce further differences...like say, a clockspeed differential ).

Even if you managed to eke out a small performance win on pixel shading vs the dedicated unit, you're still losing so much elsewhere, and it's questionable if anyone would notice your gains.

Of course all of this is speculation because those who do know aren't talking. I don't see how you're losing things elsewhere if your gpu is the limiting factor in the game though. Being able to have a idle cpu thread handle the vertex operations while being able to dedicate all your gpu processing units to pixel shading ops vs. a traditional architecture that can't use all its gpu processing units to one particular task would seem like an advantage to me.

 

[gofreak]

I don't see how you're losing things elsewhere if your gpu is the limiting factor in the game though. Being able to have a idle cpu thread handle the vertex operations while being able to dedicate all your gpu processing units to pixel shading ops vs. a traditional architecture that can't use all its gpu processing units to one particular task would seem like an advantage to me.

If there is any advantage, I don't think it'd be a big one as noted above. There may be none in fact.

You're probably losing vertex performance.

And while your game may not be CPU-bound, there's no reason why it couldn't be. You would have to make sacrifices, unless you were doing something fairly unambitious beyond graphics. Compare to a dedicated chip - its CPU could be beavering away on better physics or AI or whatever. There's a loss of potential there.

Given the goals of the Xenos design, it'd be fairly disastrous to resort to a setup where you were losing key advantages.

 

[dorio]

Given the goals of the Xenos design, it'd be fairly disastrous to resort to a setup where you were losing key advantages.

Agreed, but I don't think the goal should be to make use of platform advantages, the goal should be to make a great game even if it means not making full use of gpu efficiencies built into the system.

 

[gofreak]

Agreed, but I don't think the goal should be to make use of platform advantages, the goal should be to make a great game even if it means not making full use of gpu efficiencies built into the system.

Oh certainly, I totally agree, I was just talking purely in technical terms. The aims and goals of "a good game" and technical proficiency are not tightly bound of course (though they don't have to be mutually exclusive either).

 

[Hitokage]

Wait, I thought vertex geometry, what the Cell is supposed to help with, and vertex shading, what the RSX is supposed to have dedicated pipes for, weren't the same thing.

 

[gofreak]

Wait, I thought vertex geometry, what the Cell is supposed to help with, and vertex shading, what the RSX is supposed to have dedicated pipes for, weren't the same thing.

It could do either, but the usefulness of doing so on the CPU when there's 8 vertex shaders there is questionable in the typical case (maybe if you have REALLY heavy vertex shading requirements). Cell is more likely to be used for vertex work involving vertex creation, tesselation and such rather than "typical" vertex shader functions (animation too perhaps, which could fall under a vertex shader in some cases). And of course as part of collision detection and physics etc. all CPUs have do deal with geometry on that level (though that is not considered to be "vertex shading").

 

[dorio]

Wait, I thought vertex geometry, what the Cell is supposed to help with, and vertex shading, what the RSX is supposed to have dedicated pipes for, weren't the same thing.

Hmmmm, I thought the cell was suppose to take the place of the vertex shading pipes.

 

[gofreak]

Hmmmm, I thought the cell was suppose to take the place of the vertex shading pipes.

This was regularly speculated prior to E3, but since then there's been little discussion of it. When Tim Sweeney talks about only using the PPE with UE3 thusfar, and when vertex shading isn't amongst the tasks he mentions using the SPEs for, it's a fairly good indication that there are vertex shaders on the GPU. Of course, RSX doesn't have to match the G70 exactly in terms of configuration - they could swap some vertex shaders for pixel shaders if they wanted - but I think it's assured that there is some amount of vertex shaders on the GPU.

 

[Kleegamefan]

ATI could be really helpful here if they could just compare the performance of Xenos to, say, R480..

Unified shaders, according to them, are the way forward, but if they are going to evangelize thier advantages compared to traditional vertex/pixel architectures, why not give some performance figures compared to the X800s used in early Xenon dev systems??


Perhaps that would be too revealing???(unified shader performance, that is)

We know USAs are efficient but are they fast???

 

[gofreak]

ATI could be really helpful here if they could just compare the performance of Xenos to, say, R480..

Unified shaders, according to them, are the way forward, but if they are going to evangelize thier advantages compared to traditional vertex/pixel architectures, why not give some performance figures compared to the X800s used in early Xenon dev systems??

This would be very difficult to do. What would they use as a benchmark? PC games? Then you'll need to run on PC CPUs, and in a lot of cases they would "hide" differences between the chips (see: many benchmarks of G70 vs 6800 Ultra). They could do demos themselves, but that wouldn't be as "realworld" and there'd be complaints about their usefulness. Ditto if they just provided "paper" comparisons.

I think we'll just have to hope devs spill a little on how it is. There's an official MS doc being circulated via PM boxes everywhere about the improvements and disimprovements from alpha to beta kit with regard to the CPUs - unfortunately I haven't seen it myself yet - a similar doc about the GPUs could be interesting if one exists, I guess.

 

[Kleegamefan]

Oh you mean the rumored "Golden" PM.....yeah...I heard about it too...

 

[dorio]

ATI could be really helpful here if they could just compare the performance of Xenos to, say, R480..

Unified shaders, according to them, are the way forward, but if they are going to evangelize thier advantages compared to traditional vertex/pixel architectures, why not give some performance figures compared to the X800s used in early Xenon dev systems??


Perhaps that would be too revealing???(unified shader performance, that is)

We know USAs are efficient but are they fast???

Maybe they are waiting for ATI to show their hand with the RSX. They can't really scream superiority when they don't know what the RSX will bring to the table.

Someone needs to get a hold of that Golden document. Someone has to have it.

 

[Mrbob]

I think we'll just have to hope devs spill a little on how it is. There's an official MS doc being circulated via PM boxes everywhere about the improvements and disimprovements from alpha to beta kit with regard to the CPUs - unfortunately I haven't seen it myself yet - a similar doc about the GPUs could be interesting if one exists, I guess.

GAF has not acquired this, yet? I'm disappointed...

 

[gofreak]

GAF has not acquired this, yet? I'm disappointed...

As am I - every second poster on certain other boards seems to have it (but won't share)

 

[dorio]

As am I - every second poster on certain other boards seems to have it (but won't share)

Let's start a quest to retrieve the Golden pm. Who's with me!

 

[3rdman]

ATI could be really helpful here if they could just compare the performance of Xenos to, say, R480..

Unified shaders, according to them, are the way forward, but if they are going to evangelize thier advantages compared to traditional vertex/pixel architectures, why not give some performance figures compared to the X800s used in early Xenon dev systems??


Perhaps that would be too revealing???(unified shader performance, that is)

We know USAs are efficient but are they fast???

The only thing we've seen to demonstrate to power of Xenos thus far has been the Ruby demo. It was shown at E3 after having been ported in just 2 weeks. I believe I read a report weeks later that they had continued to optimize the demo and was able to make it run at 60 fps. (Again, I can't recall the actual website. Also, it could have been an old report based on what they showed at E3)

Not much of a benchmark, but I believe that the ruby demo shown was originally designed to showcase its R520 chip.

 

[dorio]

The only thing we've seen to demonstrate to power of Xenos thus far has been the Ruby demo. It was shown at E3 after having been ported in just 2 weeks. I believe I read a report weeks later that they had continued to optimize the demo and was able to make it run at 60 fps. (Again, I can't recall the actual website. Also, it could have been an old report based on what they showed at E3)

Not much of a benchmark, but I believe that the ruby demo shown was originally designed to showcase its R520 chip.

Never heard that about the Ruby demo running at 60 fps. I was really disappointed they they couldn't get it to run at 60fps. The demo wasn't even that impressive to me.

The only new media we've gotten that was confirmed to be from the beta kit has been from PGR3.

 

[3rdman]

I found the website and report....

http://www.computerandvideogames.com/r/?page=http://www.computerandvideogames.com/news/news_story.php(que)id=120753

However, there was no mention of 60 fps...I guess I imagined it. :/ Anyways, the article is dated 22nd of June, so at least it was running very smoothly and without hiccups. Although the action onscreen may not be much, it looks CG-like and it's very impressive but I can see why someone would be underwhelmed by it.

 

[Chiggs]

One thing to consider about this article is that when Nvidia starts making chipsets with unified shaders (and there is no "if" here, either) they'll spin it as they were the ones to finally get it "right", whereas right now, they're clamoring on about how there's still plenty of mileage with standard architecture.

I love Nvidia to death (as I won't even bother to put an ATI card in my system due to Linux issues), but what they're doing right now is similar to what ATI was trying to do with SM 3.0 last year. "Oh, it's not a big deal right nowl! There's no reason for it at this point in time! Blah, Blah, Blah!"

 

[gofreak]

One thing to consider about this article is that when Nvidia starts making chipsets with unified shaders (and there is no "if" here, either) they'll spin it as they were the ones to finally get it "right", whereas right now, they're clamoring on about how there's still plenty of mileage with standard architecture.

I love Nvidia to death (as I won't even bother to put an ATI card in my system due to Linux issues), but what they're doing right now is what ATI was trying to do with SM 3.0 last year. "Oh, it's not a big deal! There's no reason for it! Blah, Blah, Blah!"

There may be an element of that, but do you not think it also possible that it's a little premature? Just because unified shaders are right later one doesn't mean they're necessarily right now.

NVidia has learnt hard lessons about jumping the gun (see: precision in the NV3x pipeline). I'm not saying ATi have here, but it's not impossible (and let's remember this IS amongst their first SM3.0 implementations aswell as a first unified shader architecture - the risk isn't neglible).

You may be right, of course, I'm not necessarily disagreeing. But there are many sides of the coin to consider. Really all that's happening here is that NVidia and ATi are coming from 2 different perspectives with regard to the current landscape.

 

[sangreal]

I don't see why we should take David Kirk's remarks any more seriously than we take J Allard or Richard Huddy's remarks.

 

[3rdman]

There may be an element of that, but do you not think it also possible that it's a little premature? Just because unified shaders are right later one doesn't mean they're necessarily right now.

What? If its working, how can it be premature???? Thats like saying we shouldn't release gas saving engines until we run low on our oil reserves.

The benefits are obvious to the console space. Low heat, low power consumtion (expect to see this chip evolve into the mobile arena), and higher yields (supposedly). The man is speaking on behalf of Nvidia...do you really expect any kind of unbiased answer, gofreak? You know better than that.

The silliest thing about them is that in one breath they tell us that a USA isn't an avenue worth pursuing yet and eventually they will go the same route. LOL!

 

[dorio]

There may be an element of that, but do you not think it also possible that it's a little premature? Just because unified shaders are right later one doesn't mean they're necessarily right now.

I can buy that, but what do you forsee is the reason why it's not the right solution now vs. the right one in the future.

 

[Striek]

What? If its working, how can it be premature???? Thats like saying we shouldn't release gas saving engines until we run low on our oil reserves.!

Its more a matter of how its implemented. If its working, but has lower performance than a traditional GPU then of course its premature.

 

[Wunderchu]

I don't see why we should take David Kirk's remarks any more seriously than we take J Allard or Richard Huddy's remarks.

Because David Kirk is NVIDIA's Chief Scientist .. I would imagine that he has some significant involvement in the development of all hardware tech. @ NVIDIA , which likely goes beyond J. Allard's or Richard Huddy's involvement is, where hardware is concerned....

 

[Pimpwerx]

NVidia comments on ATI hw are as useful as ATI comments on NV hw. In other words, not very. :lol PEACE.

 

[sangreal]

Because David Kirk is NVIDIA's Chief Scientist .. I would imagine that he has some significant involvement in the development of all hardware tech. @ NVIDIA , which likely goes beyond J. Allard's or Richard Huddy's involvement is, where hardware is concerned....

Eh, Richard Huddy is the head of ATi's developer relations which is more pertinent to the topic at hand than the actual hardware.

NVidia comments on ATI hw are as useful as ATI comments on NV hw. In other words, not very. :lol PEACE.

exactly

 

[Chiggs]

Because David Kirk is NVIDIA's Chief Scientist .. I would imagine that he has some significant involvement in the development of all hardware tech. @ NVIDIA , which likely goes beyond J. Allard's or Richard Huddy's involvement is, where hardware is concerned....

Richard Huddy actually used to work for Nvidia, and he's actually a pretty smart guy. Not that I'm trying to contradict what you wrote, I just wanted to throw that out there.

 

[Kleegamefan]

I think we can all agree neither ATI nor nVidia are very forthcoming in the way we want them to be......

 

[3rdman]

Its more a matter of how its implemented. If its working, but has lower performance than a traditional GPU then of course its premature.

True. But, we've heard of absolutely nothing negative when it comes to Xenos...nothing except from a Nvidia rep...shocking!

 

[Mrbob]

I think we can all agree neither ATI nor nVidia are very forthcoming in the way we want them to be......

Hehe the pr postings by both companies in the last year and a half should have taught us this already! Both companies have really gone down hill in this timeframe imo.....a bunch of crap with little substance given.

 

[Wunderchu]

Hehe the pr postings by both companies in the last year and a half should have taught us this already! Both companies have really gone down hill in this timeframe imo.....a bunch of crap with little substance given.

well.. their competition is also tight right now, IMO, with neither company clearly leading the other with regards to mindshare+marketshare, AFAIK ....... so, it would not really be a good idea for either of them to actually give out any useful information, which could potentially help inform the other about their plans...

 

[mrklaw]

NVidia could go the route of redudancy to get good yields at target clock speeds on RSX.

Traditional method has been sacrifice pixel pipes for PC parts, as you still need vertex shaders. Perhaps for RSX they'd switch, and allow vertex shaders to fail and go for full pixel pipe parts?

That'd give the maximum possible pixel power, while improving yield.


Of course, I'd love them to bulldoze out the vertex shaders and stick another bunch of pixel shaders in there, but thats too much to hope..

 

[Kleegamefan]

Well there *are* some transistors on G70 that wont be required on RSX.....H.264 acceleration
for starters and god knows what else...