Why not let PowerVR build RSX?

[Simon F]

*sigh*

Every time I tell you why you're wrong, you ignore my points repeatedly. Go read this thread again, especially this post. Please stop this "TBDR destroys IMR" nonsense.

I don't see anything particular interesting in that post so I think he is justified in not taking all that much notice.

 

[Mintmaster]

I see what you mean, but even given the same time as ATI, I don't see how it's possible for nVidia to do a G80 for PS3 since it's radically different from their previous products.

I would actually argue that Xenos is more different from past ATI products than G80 is from NVidia's past products. Excluding the EDRAM, Xenos and G80 are more similar than you think.

In my understanding, ATI was on the verge of bringing out Unified Architecture, hence MS was able to capitalize on it *at the right time* for Xbox 360 first. Perhaps, ATI should be credited for its ability to come up with Xenos within a short time, rather than nVidia being rushed.

I look at it differently. Both had unified architectures simmering, but we knew about ATI's plans whereas NVidia kept it secret (thanks to Kirk's deceptive statements). Given that G80 came out way before R600, I don't think this was an issue. Both ATI and NV had the technology in the pipeline. Time and indifference were bigger factors.

I say indifference because while ATI really wanted to show technical superiority with Xenos, NVidia had G80 for that. Moreover they probably knew they were Sony's only option, and nobody's going to be running benchmarks on RSX, so it didn't make much of a difference how optimal it was.

The G7x architecture was/is the best option for PS3. Sony further added the Cell <--> RSX interaction in line with its vision (Remember there is still a SCC somewhere in the mix). I hope people who attend the RSX session can update us on its low-down after GDC.

Cell-RSX interaction is highly overrated. The CPU is there to figure out what to draw, and the GPU is there to draw it. The reason PC guys want good CPU-GPU communication is for GPGPU purposes, but that's because they don't have a Cell-type processor to work with. The idea of the GPU controlling part of the CPU like with XPS on the 360 doesn't seem particularly useful to me in terms of enabling new techniques.

 

[Fafalada]

and nobody's going to be running benchmarks on RSX

Yea I bet THAT's the reason s why Sony and NVidia spent their time making every little performance detail available from hw-performance counters (on PS2 they made us buy extra hardware to even get close to that kind of detail).

Cell-RSX interaction is highly overrated. The CPU is there to figure out what to draw, and the GPU is there to draw it.

You do realize you're contradicting yourself here?

 

[Mintmaster]

I don't see anything particular interesting in that post so I think he is justified in not taking all that much notice.

It addresses his relentless comparison of TBDR to IMR in the past. That evidence is nearly useless today because so much has changed. I know you're a big fan of TBDR too, but you cannot deny that when you consider T&L, increased vertex size and increased vertex count along with IMR improvements in culling and BW utilization that past architecture comparisons are 100% useless.

There is no way anyone can say with any certainty that today a TBDR would be a more efficient (i.e. better perf/$) solution for a console. Suggesting that the complete lack of TBDR solutions in the high end is simply due to ignorance is ridiculous.

 

[Mintmaster]

Yea I bet THAT's the reason s why Sony and NVidia spent their time making every little performance detail available from hw-performance counters (on PS2 they made us buy extra hardware to even get close to that kind of detail).

Sorry, I didn't write that very clearly.

I meant comparison benchmarks. You're not going to see geeks saying NVidia sux because ATI is 10% faster in some game or something, because it's a closed console. About the closest thing you could do is port a test to XB360, but that's not going to happen very often. The point is that once they had the contract, competition was gone. The impetus to try designing something novel and amazing wasn't there.

(BTW, they didn't spend any time on implementing those counters. NVidia and ATI have been doing that for generations to help themselves with future design decisions.)

You do realize you're contradicting yourself here?

Not in the context patsu is suggesting, namely this "vision" where the modification to G7x concerning Cell-RSX communication does wonderous things.

 

[nAo]

Cell-RSX interaction is highly overrated. The CPU is there to figure out what to draw, and the GPU is there to draw it.

Straighforward ideas are always the ones that are likely to be implemented first, exotic stuff naturally comes later.

 

[Simon F]

It addresses his relentless comparison of TBDR to IMR in the past. That evidence is nearly useless today because so much has changed. I know you're a big fan of TBDR too, but you cannot deny that when you consider T&L, increased vertex size and increased vertex count along with IMR improvements in culling and BW utilization that past architecture comparisons are 100% useless.

There is no way anyone can say with any certainty that today a TBDR would be a more efficient (i.e. better perf/$) solution for a console. Suggesting that the complete lack of TBDR solutions in the high end is simply due to ignorance is ridiculous.

A "big fan" you say? Could it be that I might know a little bit more about the workings of TBDRs than you do?

 

[Techno+]

There is a interesting slide Imagination tech showed once

http://www.tomshardware.com/picturegalleries/gallery-200701251-5.html

 

[Xmas]

Ps3 may have set specs, but it's a hell of a lot closer to a PC than it is to a cell phone. Your argument would be pretty good for say...PSP2 (though PSP already looks like it outdid powervr's handheld graphics from a company with no real experience in the field),

Comparing the PSP GPU to MBX is comparing apples to bananas. Although they're both used for handheld graphics they were created with entirely different design goals: power consumption, die area, cost, lifetime, design timeframe, scalability, number of integrators and platforms, feature set, standards conformance and software requirements are all different.

 

[Megadrive1988]

I would actually argue that Xenos is more different from past ATI products than G80 is from NVidia's past products. Excluding the EDRAM, Xenos and G80 are more similar than you think.

I look at it differently. Both had unified architectures simmering, but we knew about ATI's plans whereas NVidia kept it secret (thanks to Kirk's deceptive statements). Given that G80 came out way before R600, I don't think this was an issue. Both ATI and NV had the technology in the pipeline. Time and indifference were bigger factors.

I say indifference because while ATI really wanted to show technical superiority with Xenos, NVidia had G80 for that. Moreover they probably knew they were Sony's only option, and nobody's going to be running benchmarks on RSX, so it didn't make much of a difference how optimal it was.

Cell-RSX interaction is highly overrated. The CPU is there to figure out what to draw, and the GPU is there to draw it. The reason PC guys want good CPU-GPU communication is for GPGPU purposes, but that's because they don't have a Cell-type processor to work with. The idea of the GPU controlling part of the CPU like with XPS on the 360 doesn't seem particularly useful to me in terms of enabling new techniques.

very much agreed, Mintmaster. While RSX has been optimized to work with CELL,
the CELL>< RSX 'special' co-operation /interaction isn't really different than the typical console CPU><GPU co-operation.

It baffles me why SCEI did not ask Nvidia to co-design a G8x based GPU with eDRAM, since G80 had been in development since 2002, before the GeForce FX was released. even though a full blown GF8800 GTX would be to costly for a console, something like GTS or a little even a somewhat, but combined with eDRAM l would've been so much better than the NV47-derived RSX which is just plain outdated, though still modern compared to the Graphics Synthesizer family.

 

[patsu]

I would actually argue that Xenos is more different from past ATI products than G80 is from NVidia's past products. Excluding the EDRAM, Xenos and G80 are more similar than you think.

All I'm saying is: ATI may be *earlier* with UA and they used the opportunity to roll it out on Xbox 360.

The EDRAM part is where the customization (for Xbox 360) is [QUESTION: So the IP is owned by MS now ?]. Without that part, Xenos looks like a natural progression from previous thinking. G80 seems more radical in thinking to me (because of its switch to go scalar).

I look at it differently. Both had unified architectures simmering, but we knew about ATI's plans whereas NVidia kept it secret (thanks to Kirk's deceptive statements). Given that G80 came out way before R600, I don't think this was an issue. Both ATI and NV had the technology in the pipeline. Time and indifference were bigger factors.

Yes, in my view timing is the difference. ATI has an early shot at a UA GPU before nVidia, matching Xbox 360's timeline, and they used it on Xenos. nVidia may have preferred to use it on their PC part to sustain the lead. I can't fathom them rolling out a G8x driver one year early (for PS3 devs to use) given their current driver situation.

I say indifference because while ATI really wanted to show technical superiority with Xenos, NVidia had G80 for that. Moreover they probably knew they were Sony's only option, and nobody's going to be running benchmarks on RSX, so it didn't make much of a difference how optimal it was.

It's possible, but this does not support the notion that RSX was a last minute decision or rushed. In fact, if they go G80, then I'd say the GPU and SDK will be even more rushed.

Cell-RSX interaction is highly overrated. The CPU is there to figure out what to draw, and the GPU is there to draw it. The reason PC guys want good CPU-GPU communication is for GPGPU purposes, but that's because they don't have a Cell-type processor to work with. The idea of the GPU controlling part of the CPU like with XPS on the 360 doesn't seem particularly useful to me in terms of enabling new techniques.

I'll reserve my judgement until more is known about RSX and Cell interaction. Devs have been dropping hints like:

Straighforward ideas are always the ones that are likely to be implemented first, exotic stuff naturally comes later.

 

[Mintmaster]

A "big fan" you say? Could it be that I might know a little bit more about the workings of TBDRs than you do?

Chill out, I didn't mean that in a derogatory way (I would have put three asterisks after the word fan if I did). Should I have said proponent instead? I know you're very knowledgeable, and was disappointed not to get input from both you and Kristof in the other TBDR thread.

Besides, wouldn't it be more constructive if you countered my arguments instead of implying I'm talking out of my ass?

 

[Carl B]

Not in the context Carl B is suggesting, namely this "vision" where the modification to G7x concerning Cell-RSX communication does wonderous things.

Speaking of context... the only sentence in this thread I used the words 'vision' and RSX together was one in which I highlighted a lack of said vision:

Which is to say, that if there are shortfalls in the design of RSX, the shortfalls are of vision...

So, I'd appreciate if you didn't weight my words incorrectly when refering to my past posts as the source of your arguments.

I didnt say that the Cell-RSX communication was going to do 'wondrous things' either; what I said was that there is a material consideration to at least acknowlege when refering to RSX as a straight G7 port, because there are a few noteworthy changes on the hardware level beyond the desktop parts.

I'm personally of the mind that ultimately dev API familiarity and shader flexibility is what led Sony down such a 'conventional' path with RSX. Conventional in quotes because, well, we really should note that there are important system considerations in place, however minor on a hardware level they may be.

 

[Mintmaster]

All I'm saying is: ATI may be *earlier* with UA and they used the opportunity to roll it out on Xbox 360.

Yes, in my view timing is the difference. ATI has an early shot at a UA GPU before nVidia, matching Xbox 360's timeline, and they used it on Xenos.

I said time, not timing. IMO both had the technology at nearly the same time.

The EDRAM part is where the customization (for Xbox 360) is [QUESTION: So the IP owned by MS now ?]. Without that part, Xenos looks like a natural progression from previous thinking. G80 seems more radical in thinking to me (because of its switch to go scalar).

The decision to go scalar is much less radical than you think. Xenos has a pixel grouper that gets 16 quads (possibly from different triangles but with the same renderstate) and puts them in a "vector". It then assigns one of the three ALU groups to work on that vector. The ALU groups rapidly switch working on many vectors to hide texture latency.

As it is, there are 16 vec4+scalar processors in each ALU group, so it takes 4 cycles to get through one instruction on one vector. All ATI would have to do is change the ALU groups to instead operate on one channel of all 64 pixels in the vector every cycle.

Beyond the DX10 features and eDRAM, it's actually quite surprising how similar G80 and Xenos are when you look at the data flow diagrams.

I'll reserve my judgement until more is known about RSX and Cell interaction. Devs have been dropping hints like:

Well that particular hint is a platitude applicable to every aspect of life (no offense, nAo :smile: ). I have yet to see or hear about anything new this enables beyond making procedural geometry a bit more efficient or slightly mitigating the disadvantages of a split memory pool.

very much agreed, Simon. While RSX has been optimized to work with CELL,
the CELL>< RSX 'special' co-operation /interaction isn't really different than the typical console CPU><GPU co-operation.

I think Simon might resent you confusing me with him.

 

[Mintmaster]

Speaking of context... the only sentence in this thread I used the words 'vision' and RSX together was one in which I highlighted a lack of said vision:

Oops. I meant patsu, not you. Sorry about that.

 

[patsu]

I said time, not timing. IMO both had the technology at nearly the same time.

Ok... but then why is R600 delayed again ? Having the technology and rolling out a good product are 2 separate things. ATI would be further along in making/prototyping some form of Xenos at that time.

The decision to go scalar is much less radical than you think. Xenos has a pixel grouper that gets 16 quads (possibly from different triangles but with the same renderstate) and puts them in a "vector". It then assigns one of the three ALU groups to work on that vector. The ALU groups rapidly switch working on many vectors to hide texture latency.

As it is, there are 16 vec4+scalar processors in each ALU group, so it takes 4 cycles to get through one instruction on one vector. All ATI would have to do is change the ALU groups to instead operate on one channel of all 64 pixels in the vector every cycle.

Beyond the DX10 features and eDRAM, it's actually quite surprising how similar G80 and Xenos are when you look at the data flow diagrams.

Ultimately, nVidia went scalar and ATI did not. It's the change in thinking that is bold to me. Many interesting ideas look simple after crossing the initial thought barrier. It is unknown to me how long it took them to experiment, profile and finally decide to go G80 route. Does it also mean that they have to redo their internal software from scratch ?

Well that particular hint is a platitude applicable to every aspect of life (no offense, nAo :smile: ). I have yet to see or hear about anything new this enables beyond making procedural geometry a bit more efficient or slightly mitigating the disadvantages of a split memory pool.

Sure... but there'd be no reason to post if nAo doesn't have any to add ? Do you have RSX documentation ?

 

[Lazy8s]

Aurora's been in use for a while. Although Sega Sammy may build it into everything from low cost arcade boards to handhelds -- http://gakusei.enjapan.com/08/images/00/05/000541_cdocVIDEO_000003368_V01.wmv , it's been made as their platform for the new pachislot format that the industry is mandated to adopt.

Since some manufacturers missed adopting PowerVR back when its advantages were so pronounced, their understanding of its architecture is obviously suspect and potentially the chief reason they're not using it now, regardless of whether or not it still has the edge technologically.

 

[Mintmaster]

Ok... but then why is R600 delayed again ? Having the technology and rolling out a good product are 2 separate things. ATI would be further along in making/prototyping some form of Xenos at that time.

Okay, I'll agree with you there, but relative to console launch times, G80 came out the same time as Xenos. Even if Sony intended for PS3 to be released earlier, there's only a few months difference we're talking about. Is that enough to impact design choices? Maybe, but I doubt it.

I just don't agree with the idea that 30 months before the 360 launch ATI had better unified shader tech than NVidia did 36 months before the PS3 launch. I do, however, think that ATI working with MS 16 months before NVidia worked with Sony impacted design choices. Whatever arguments you make about NVidia working with Sony before the press release can be applied to ATI and MS.

Anyway, I think we've beaten this topic to death.

Ultimately, nVidia went scalar and ATI did not. It's the change in thinking that is bold to me.

Okay, I'll give you that. I thought we were talking about how different the architectures were from previous stuff, that's all. Sure, going scalar changes how you compile shaders, but it should get easier, not harder.

My point is that R4xx --> Xenos is much more radical to ATI than G7x --> G80 for NVidia, but you claimed the opposite.

Sure... but there'd be no reason to post if nAo doesn't have any to add ? Do you have RSX documentation ?

I have some, but not nearly as much as someone with an SDK, obviously. I was just hoping nAo would give something a little more concrete. I got lots of wild ideas about how to use branching, vertex textures, pixel shaders, and EMBM in novel ways when these features came out, and read lots of papers with cool ideas which don't get used in games for many years. However, nothing at all comes to mind here beyond what I mentioned before. AGP and PCI-E have generally had bandwidth close to system bandwidth, so it's expected for Flex I/O to offer transfer speeds comparable to XDR's bandwidth. Procedural geometry isn't particularly exciting to me, especially since you can do it (at some XDR BW cost) without any XPS-type interaction.

I would very much love for nAo or some other dev to show me otherwise.

[BTW is there any name for RSX's ability to control SPUs? XPS - XBox Procedural Synthesis - is the analogue on the 360, AFAIK. Again, if anyone knows more, please share.]

 

[patsu]

I have some, but not nearly as much as someone with an SDK, obviously. I was just hoping nAo would give something a little more concrete. I got lots of wild ideas about how to use branching, vertex textures, pixel shaders, and EMBM in novel ways when these features came out, and read lots of papers with cool ideas which don't get used in games for many years. However, nothing at all comes to mind here beyond what I mentioned before. AGP and PCI-E have generally had bandwidth close to system bandwidth, so it's expected for Flex I/O to offer transfer speeds comparable to XDR's bandwidth. Procedural geometry isn't particularly exciting to me, especially since you can do it (at some XDR BW cost) without any XPS-type interaction.

I would very much love for nAo or some other dev to show me otherwise.

[BTW is there any name for RSX's ability to control SPUs? XPS - XBox Procedural Synthesis - is the analogue on the 360, AFAIK. Again, if anyone knows more, please share.]

Aha ! Now we are talking... hmm... how shall I put it ? I'll be blunt. Can I have a lookie lookie at your RSX material ? pretty please ?

 

[rounin]

I would very much love for nAo or some other dev to show me otherwise.

[BTW is there any name for RSX's ability to control SPUs? XPS - XBox Procedural Synthesis - is the analogue on the 360, AFAIK. Again, if anyone knows more, please share.]

Therein lies the problem - nAo is at most as obliged to share with you as you are to share with us your documentation. And even then, once he shares with you, there is no point since it would have to stop at you anyway and the only way we would know about it would be if you just stopped talking about this point...at which point it would be obvious that nAo has talked to you and then....

the Sony Ninjas will kill you both.