I don't know why you are so confident that the reason that Nvidia hadn't introduced a US-based architecture because they didn't know how as opposed to it being a strategic decision based on what architecture would be the most performant when running the games that consumers would be playing during its primary product life. That decision is markedly different when creating a product that will be replaced in 12-24 months vs. one that will need to be viable for 60+. Also, despite being the first to offer a US design with Xenos, ATI were actually later introducing their first US-based part for the PC market (Radeon HD 2900, May 14th 2007) than Nvidia (Geforce 8800 GTX, November 8, 2006).
PS3 hardware design choices - good or bad? *spawn
- Thread starter DoctorFouad
- Start date
I don't know why you are so confident that the reason that Nvidia hadn't introduced a US-based architecture because they didn't know how as opposed to it being a strategic decision based on what architecture would be the most performant when running the games that consumers would be playing during its primary product life. That decision is markedly different when creating a product that will be replaced in 12-24 months vs. one that will need to be viable for 60+. Also, despite being the first to offer a US design with Xenos, ATI were actually later introducing their first US-based part for the PC market (Radeon HD 2900, May 14th 2007) than Nvidia (Geforce 8800 GTX, November 8, 2006).
I don't understand. US GPUs don't need game to be written for them to benefit. What are the advantages of G70 over a theoretical US part released that year that makes G70 better, and why would those benefits of discrete shaders not be aplpicable to a console? What changed in 2006 in the PC space so that nVidia's choice was to move from discrete shaders to unified, where a change to US in 2005 would have been a bad choice?
I think Kutaragi said something about using 8 SPUs because 8 was a beautiful number in computing. I hope he was joking and that they had better reasons than that.
I suspect there is more than a little truth in that.
Alternatively they could have put that extra power towards the GPU and clocked it at 550mhz which ought to have yielded an efficiency boost as well. Beyond anything else it would have made the system cheaper and more powerful out of the gate.
Is there a large benefit to implementing more complex shader units with the flexibility to adapt to different pixel/vertex loads when you're almost exclusively going to be running games that are optimized to run on hardware with a fixed ratio of vetex to pixel shader units? Given a target transistor count and die size would you expect this design to show a performance benefit running those games over one that could potentially contain more of the simpler fixed-function cores.
Vista came out and ushered in DX10 which, while it didn't technically require a US architecture, had new requirements that made a US architecture the best choice. How were Nvidia and ATI to know both Vista and DX10 would flop? Even if Vista didn't prompt a move to US, though, 360->PC ports would have.
I've said it once and I'll say it again, for the last 12-18 months PS3 has been kicking out much more impressive visuals in my eyes so it's already holding it's own very nicely.
Regardless of your personal opinion about exclusive titles which there's very little point arguing about, the ps3 continues to lose most head to heads in third party titles. If it had gone the big gpu and normal cpu route like I suggest above it would be winning them.
Third party titles don't make a system though do they, exclusive ones do.
And RSX+Cell combo don't seem to be holding exclusive games back so if third party games are struggling I would put that down the developers and not the machine.
Exclusive first party games has always and always will be the only way to judge a machines muscle as third party games would never be able to harness the machine as much as there's too many factors/variables in third party games.
So simply wanting a bigger GPU+Normal CPU to match up in third party titles seems silly and wasteful.
Well, that answers the question of RSX then. If it's true that fixed-function units give better utilisation (which it isn't) then RSX had fixed function units to which games could be highly optimised unlike the same component in a PC (where US would be more beneficial). If it's not true, and having the driver dynamically and transparently adjust workloads according to requirements across US, then nVidia should have had such a part in the PC space ASAP because it'd be most beneficial there. Either way, your premise is disproven.
You can try to target that 1:3 ratio but it's a pretty unrealistic goal. Especially when you consider how much frame time a modern game will dedicate to post processing passes, where your vertex shader units just sit around idling.
Well I believe that the gpu is not the only issue of the design.
There are two memory pools. Gpu is lacking bandwidth. The system uses more expansive ram with not neat benefits.
The whole project has imo be undermined by a lack of a clear design idea.
The cell shows the limits of design by committee. Sony wanted vectors units, it's not pushing to say that they wanted successor for the vpu 1&2 of the ps2.
Toshiba wanted mips cores, Ibm power core. It ended up with a compromise and a isa not that graphic friendly. It turns out this way because Sony alone could not fund the project and possibly felt like it didn't had the competences to do it alone.
So for Sony the most important part the one there was the most closely involved into.
Once i.the project started running and contracts signed as well as deadlines more or less known, Sony was trapped. I believe that soon enough Sony realized that would not get the EE successor they were expecting and hence contacted Nvidia. I believe that put from scratch put a dent in the ps3 profitability plan (which already had to take in account brd expenses). Nv is not the kind to let anything out of its doors for a bargain. I believe that the 10 years lifespan was a cover up for that matter of fact, not something envisioned in early design stage.
Overall my belief is that the project was derailed as soon as Sony got involved into the STI alliance. Lot of resources were to be thrown at the cpu side for something that could not be the specialized cpu that Sony would have needed in a try to push EE / GS philosophy a further.
By looking at the Cell (and what we know about their requirement for the project) I will never believe otherwise.
The only winner in this story were imho IBM and Rambus.
There are two memory pools. Gpu is lacking bandwidth. The system uses more expansive ram with not neat benefits.
The whole project has imo be undermined by a lack of a clear design idea.
The cell shows the limits of design by committee. Sony wanted vectors units, it's not pushing to say that they wanted successor for the vpu 1&2 of the ps2.
Toshiba wanted mips cores, Ibm power core. It ended up with a compromise and a isa not that graphic friendly. It turns out this way because Sony alone could not fund the project and possibly felt like it didn't had the competences to do it alone.
So for Sony the most important part the one there was the most closely involved into.
Once i.the project started running and contracts signed as well as deadlines more or less known, Sony was trapped. I believe that soon enough Sony realized that would not get the EE successor they were expecting and hence contacted Nvidia. I believe that put from scratch put a dent in the ps3 profitability plan (which already had to take in account brd expenses). Nv is not the kind to let anything out of its doors for a bargain. I believe that the 10 years lifespan was a cover up for that matter of fact, not something envisioned in early design stage.
Overall my belief is that the project was derailed as soon as Sony got involved into the STI alliance. Lot of resources were to be thrown at the cpu side for something that could not be the specialized cpu that Sony would have needed in a try to push EE / GS philosophy a further.
By looking at the Cell (and what we know about their requirement for the project) I will never believe otherwise.
The only winner in this story were imho IBM and Rambus.
Last edited by a moderator:
Well, that answers the question of RSX then. If it's true that fixed-function units give better utilisation (which it isn't) then RSX had fixed function units to which games could be highly optimised unlike the same component in a PC (where US would be more beneficial). If it's not true, and having the driver dynamically and transparently adjust workloads according to requirements across US, then nVidia should have had such a part in the PC space ASAP because it'd be most beneficial there. Either way, your premise is disproven.
You didn't address my scenario where simpler functional units could allow for a greater number of functional units within a given transistor budget. G71 was 278M transistors, G80 was 686M counting the 5M for the NV I/O core. How do you think a G71 derivative with the same number of transistors and the same memory bandwidth of the G80 might have performed in game benchmarks at the time the G80 was released?
Also, as I mentioned, you're basing your whole assumption that Nvidia didn't have the capability to produce a US-based architecture prior to G80 on them not producing one for the PC space before then and nothing else. Yet, as I mentioned, it took ATI (who had already proven that they could produce a US-based architecture) even longer to produce a US-based architecture for the PC space. So prove your premise.
Which have shortcomings. The console space graphics are exactly the same, and so have exactly the same requirements, as the PC space; perhaps there's an argument that fixed hardware allows for better low-level balancing in the engine design, but the move to middelwares has reduced that. So, whatever the best solution for the PC is would also be the best solution for the PS3. If unified shaders are better, than nVidia should have been putting US into the PC. If less-flexible cores is better, then nVidia should have been putting more and less flexible cores into the PC (G70).
That's how we know nVidia could not produce a US part earlier - because if they could, they would have. US is better (when executed properly). nVidia went on record around the time of Xenos to point out US's 'shortcomings' and tell everyone that fixed-function was better in a console. We all know how that panned out. And RSX isn't imbalanced because they chopped off a couple of vertex units or anything. It just happens that the workloads devs wanted ended up doing more vertex work, which Xenos could adjust too (plus I don't think RSX's vertex pipes were particularly strong by design, but I don't claim any great knowledge there). RSX was basically nVidia's flagship architecture, doing the best nVidia could do for rendering DX9 graphics. If nVidia were capable of better but chose not to, more fool them!
As for ATi, I'm guessing R600 was delayed so they could refine the approach as they weren't 100% happy with the first-gen US and the cost of rolling out a new part within the context of their roadmap, along with whatever glitches and bugs set them back. The experience of Xenos would have been very beneficial to refining their approach. nVidia, without that previous experience, could only test their first US designs in the PC.
G80 had 2.5x the transistors but on average G80 was 2.5x as fast and up to 3-4x as fast as G71 in certain situations so I would imagine that G71 with 2.5x transistor budget would still be slower.
G80 does all that and still only consumes 40w more power on average over a 7900GTX.
I don't believe that unified vs specialized vertex and pixel units tell the whole story when comparing the PS3 and the 360.
I believe that Edram did (and still do) more for the 360 than the ALUs differences. Way more.
I also believe that the extra RAM and more flexibly allocated than the two the ps3 has has does a lot more for the 360 than the Xenos unified shaders ALUs.
RSX main weakness is bandwidth. Overall I would say that the ps3 suffered more from its 2 memory pool and the pretty low bandwidth (vs xenon not bad by self at the time of shipping) for frame-buffer operation than from RSX inner architecture.
Even with weaker vertex through put I remember Nao stating that there were ways to deal with it without involving extra SPU work. / I would summup as the issue is way overblown.
Looking at the Ps2 and the lot of bandwidth the GS (eventhough I read here that it was not that efficient and less cleverly used could have done the trick) had to play with the ps3 looks to me like half a design as the Cell tried to be EE2 (and way to more due to multiple design constrain) and nothing looks like the matching GS in the design.
Another weakness of the design is the 2 128bit buses which possibly prevent Sony to comes with a Vahalla like chip (either constrain the size or the mem controller constrain the power dissipation).
For me a G80 was not the matching solution to what seemed Sony goals (and was not doable anyway, which is not to say that anything else was impossible).
A RSX with a daughter die ala Xenon would have been way better without touching the ALU. It would have been less costly to develop than a new architecture (personal bet but I think the odd are that I'm right).
Another issue of the system is the link between the Cell and the Rsx which doesn't seem to offer the promise bandwidth. May it had worked properly Sony may have go for a single pool of memory (XDRAM) with the GPU accessing the RAM through the cell and framebuffer operation would have been handle on a daughter die (including EDRAM).
Overall it looks like a patchwork where (costly) chips have been designed a bit too much in insulation. There are many lacking to the design in my opinion discussing RSX alu's only is a short sighted argument (not try to sound rude but I see other way to express what I think).
I believe that Edram did (and still do) more for the 360 than the ALUs differences. Way more.
I also believe that the extra RAM and more flexibly allocated than the two the ps3 has has does a lot more for the 360 than the Xenos unified shaders ALUs.
RSX main weakness is bandwidth. Overall I would say that the ps3 suffered more from its 2 memory pool and the pretty low bandwidth (vs xenon not bad by self at the time of shipping) for frame-buffer operation than from RSX inner architecture.
Even with weaker vertex through put I remember Nao stating that there were ways to deal with it without involving extra SPU work. / I would summup as the issue is way overblown.
Looking at the Ps2 and the lot of bandwidth the GS (eventhough I read here that it was not that efficient and less cleverly used could have done the trick) had to play with the ps3 looks to me like half a design as the Cell tried to be EE2 (and way to more due to multiple design constrain) and nothing looks like the matching GS in the design.
Another weakness of the design is the 2 128bit buses which possibly prevent Sony to comes with a Vahalla like chip (either constrain the size or the mem controller constrain the power dissipation).
For me a G80 was not the matching solution to what seemed Sony goals (and was not doable anyway, which is not to say that anything else was impossible).
A RSX with a daughter die ala Xenon would have been way better without touching the ALU. It would have been less costly to develop than a new architecture (personal bet but I think the odd are that I'm right).
Another issue of the system is the link between the Cell and the Rsx which doesn't seem to offer the promise bandwidth. May it had worked properly Sony may have go for a single pool of memory (XDRAM) with the GPU accessing the RAM through the cell and framebuffer operation would have been handle on a daughter die (including EDRAM).
Overall it looks like a patchwork where (costly) chips have been designed a bit too much in insulation. There are many lacking to the design in my opinion discussing RSX alu's only is a short sighted argument (not try to sound rude but I see other way to express what I think).
I can agree in part. Us was essential to Xenos to get the same performance as more shaders in less transistors, leaving part of the budget free for eDRAM. If Xenos had fixed shaders, it'd be conisderably underpowered compared to RSX. That said, some of the typical shortcomings in PS3 games are due to lack of RAM, thanks to Sony's lack of finesse with the OS and I guess less efficient use of RAM due to split pools. There are some bandwidth limits leading to lower resolutions, but there are also fill limits I think.
It'd perhaps be poignant to the discussion to have a list of PS3's graphical shortcomings to identify what could have been improved in the GPU and overall system. I do believe in some cases Cell could be used to generate effects much more efficiently than their current GPU implementations, but that'd need a developer to sink research into these ideas why few can afford to do while creating actual games, when they are already developing tech based on existing research papers and ideas.
It'd perhaps be poignant to the discussion to have a list of PS3's graphical shortcomings to identify what could have been improved in the GPU and overall system. I do believe in some cases Cell could be used to generate effects much more efficiently than their current GPU implementations, but that'd need a developer to sink research into these ideas why few can afford to do while creating actual games, when they are already developing tech based on existing research papers and ideas.
I honestly don't think it's that clear cut. G80 included more architectural innovations for Nvidia than just the move to US, some of which were required for DX10 compliance, others not (split clock domains come to mind). I don't agree that the timing of the launch of G80 gives any definitive indication that any one part of that architecture couldn't have been implemented earlier. In fact, as I keep pointing out, I don't think G80's launch is any more an indication of when Nvidia could have produced a US-based console part than Xenon's launch turned out to be an indication of when ATI were going to be ready to lauch a US-based PC part.
OTOH, I am coming around to the belief that Xenos was NOT what either Sony or Nvidia were expecting when plans for RSX were formulated and that they probably thought RSX as conceived was going to be at least the equal of Xenos when it launched. In that light, perhaps the criticisms of Sony for choosing RSX as their GPU are a bit unfair and more praise is due to ATI and MS for being super-aggressive and (hardware reliability aside) succeeding. If they hadn't been able to launch first, maybe the increase in vertex loads in multiplatform titles that exposed the weakness in RSX wouldn't have happened.
Gamesindustry.biz has run a two-part piece by former SCEE executive Richard Browne entitled, "The Rise and Fall of Sony". Part 2 of this article covers the period around the launch of the PS3 through the lauch of the Vita and the present day and ends with his thoughts on how Sony can best be successful going forward.
His narrative of the eventual PS3 design as a deviation from the what was initially conceived that was made necessary by an accelerated launch timetable seems very plausible to me. It explains the sense that I have always had that the PS3's components don't form a cohesive whole in the same way that those of their previous consoles (PS2's most especially) do.
His narrative of the eventual PS3 design as a deviation from the what was initially conceived that was made necessary by an accelerated launch timetable seems very plausible to me. It explains the sense that I have always had that the PS3's components don't form a cohesive whole in the same way that those of their previous consoles (PS2's most especially) do.
Similar threads
- Locked
