Charlie/Groo claimed some days ago that Sony may be evaluating for the PS4 some chip based on the Emotion Engine. The rumour sounds pretty much unlikely, it's would make more sense if the chip were indeed intended for the PSP2 but that's actually not my point
It got me thinking about how a PS3 could have look based on an evolution of the EE GS combo.
Here a die shot of the last EE/GS revision:
The chip is approximately 60mm² on a 90nm process and ~55 millions of transistors.
The "game" could be to wonder about what Sony could have done with say ~5 times the silicon budget ~300 millions and to guesstimate how far or close the system may have end from the PS3 we all know in regard to overall performances while maintaining backward compatibility.
It's been a while since I last did PS2 programming so let's see...
What kind of hardware evolution do you have in mind for this hypothetical EE+ chip? Let's say you replace the VPUs with SPUs so you what you end up is a Cell with a MIPS isa instead of PowerPC. That would certainly be useful.
Now for the GS that would need a major overhaul. There's only so much you can do with massive overdraw so you could probably extend the 4MiB EDRAM to something more reasonable for HD content and you also want to have some shader capabilities as well, basically Sony's version of Xenos
Could have been a very interesting piece of hardware had they gone that way, and even nice to work with if DMA setup was not PITA like in the PS2. I seriously doubt 300 million transistors would get you performance anywhere near the actual PS3, though.
and to guesstimate how far or close the system may have end from the PS3 we all know in regard to overall performances while maintaining backward compatibility.
Cell is already an obvious and rather direct evolution of EE, and outside of fine tweaking (add a few instructions to SPE ISA or maybe switch PPC for Mips core), there's not much more you can do with it.
Likewise goes for the rather fat interconnect between CPU and GPU - so the only thing you can really make more PS2-like would have been the GPU itself. And if I was a betting man, I'd say that would have been something with an order of magnitude (at least) more Fillrate then current PS3/360, and obviously a lot more eDram as well.
Something like an updated GS32 from GScube clocked at 1.6 GHz. So it would be Cell plus that. It would be underpower compare to what PS3 is, but it probably wouldn't break BC.
It's been a while since I last did PS2 programming so let's see...
What kind of hardware evolution do you have in mind for this hypothetical EE+ chip? Let's say you replace the VPUs with SPUs so you what you end up is a Cell with a MIPS isa instead of PowerPC. That would certainly be useful.
Well I was thinking about it, actually putting this post together took a hell lot of time.
For reference 5 times the PS2 is:
5 MIPS CPU
10 vectors units
80 pixel pipelines
~20 MB of on-board memory.
The split (CPU/GPU) in die area was roughly 50/50 notwithstanding the huge disparity in transistors count between the GPU and the CPU.
I was considering a lower power profile than the PS360 EE V.2 :
@600MHz
2 MIPS CPUs cores
512 kB of L2 cache
1 super "Vu1" unit
*CPU core (x2):
2 issue
32KB of instruction cache
16KB of data cache
128KB of scratch pad memory
***Vu0 (x2):
integrated to the core, think altivec.vmx/SSe unit
=> direct access to the 64KB of scratch pad memory
16 wide
Linked to the "super Vu1" unit
DMA access to the main ram and the on-board memory pool (12MB see below)
*Super Vu1 unit (x1)
64 wide
256KB of scratch pad memory
Linked to the GS V2.
DMA access to the main ram and the on-board memory pool (12MB see below)
GS V.2:
Now for the GS that would need a major overhaul. There's only so much you can do with massive overdraw so you could probably extend the 4MiB EDRAM to something more reasonable for HD content and you also want to have some shader capabilities as well, basically Sony's version of Xenos
As you say it's unclear where a "dumb" multiplication of the resources along with improvements and additions while keeping the fixed function nature would have gotten them. Still fixed function hardware has its advantage it can be quiet faster than programmable so it could have been interesting.
I was searching for when Ms introduced shaders (back to directx8) and found this presentation, it's pretty dated but the results are still interesting it reminds how costly can be pixel shading.
The Graphics Synthesizer 2 could have been the last and greatest avatar of now extinct lineage.
Actually I came to think that Sony outside of OpenGL/Directx standard could have bent the pre-directx8 fixed function pipeline into something both more flexible and interesting
it's possible that what I'll say is not possible or doesn't make that much sense I speak about as much as 40 design wins...
:
Sony could have:
*Create fixed function pipelines design to handle deferred renderer.
*up-to-date texture units.
*Hardware Support for some clever texture/render target format like RGBM.
*Have some pixel pipelines optimized for forward rendering, they would handle transparent objects and would take care of of backward compatibility. Could dedicated fixed function hardware of example in-order transparencies at a reasonable cost?
*Sony may have plan for MLAA from scratch, they could have the ROP/RBE optimized just to save bandwidth, why not some fixed function "shadow filtering" hardware?
*While keeping the pipeline fixed function Sony may have open it a bit, think memexport. Results may be output to EDRAM or RAM at some critical stage of the pipeline. So devs would not see a pipeline but a pool of resources even if fixed function/not programmable. Actually to export results could be reworked by the "super Vu1 unit"
which could access as I said both the main ram and e-dram through DMA access, and is a pretty potent vector unit (64 wide)
whom is programmable, new results could be submitting at some points of the fixed function "pipeline" say for sampling in tex units for example thus introducing some flexibility in the fixed function pipeline.
I think that Sony could have fit something pretty potent with the aforementioned silicon budget, for reference a radeon 9700 was around 100 millions transistor and that include:
vertex shader (handle via Vuo units and most likely super Vu1 unit), the "video engine",ROP/RBE design to handle MSAA, and all the transistors associated to directx/opengl support.
Say @400MHz Sony could have pack quiet some power in a pretty small die size area comparatively to RSX & Xenos.
MISC:
*Rework DMA engine
*memory interface: 128bit bus
*Main RAM size: 512MB (actually why not 1024MB? Sony would have done quiet some saving elsewhere and that would have given the system quiet an edge).
*12MB of EDRAM way more flexible than the Xenos EDRAM, accessible by GPU, CPUs, SVu1.
*Tiny CPU used for security and background task (os, sound, network, etc.)
*Good scaler
*Hardware decoder for bluray
Could have been a very interesting piece of hardware had they gone that way, and even nice to work with if DMA set-up was not PITA like in the PS2. I seriously doubt 300 million transistors would get you performance anywhere near the actual PS3, though.
Well while different I think that the system could have hold on its own and that the perceived difference may have been minimal from the "average Joe's" POV
if the description I made of the system makes any sense which given my knowledge on the matter is doubtful
.
By using raw numbers:
The system would have delivered x19 (~) the geometry throughput of the ps3:
PS2: ~6GFLOPS (I put the FPU aside...)
(VU0 2.44 / VU1 3.08GFLOPS)
Say the vector units keep mostly the same characteristic:
PS3: ~138GFLOPS
2 Vo0 => 2.44 x2 (2 units) x2 (twice the clock speed) x4(four times as wide) ~39GFLOPS
1 SVu1 => 3.08 x2(clock speed) x16 (16 times as wide) ~99GFLOPS
For reference 5 SPU are worth 25.6x5=128 GFLOPS, 6 SPU ~154GFLOPS. So the EE/GS would not be that anaemic keeping in mind that vertex workload would be more taxing on resources than on the PS3 we know (no vertex pipeline).
Then come the GPU and there is no easy figures to provide but I feel like a properly design fixed function pixel pipeline on top of EDRAM may have ended surprising more than one person.
Overall the system may have been obviously less powerful than the Cell/RSX but may have suffered from less communication overhead a great leveller could have been main RAM, in regard to the saving made Sony could have consider 768 or 1024MB
I'll echo the sentiment that the SPEs function like generalized VUs, with the ability to source and produce data from/to anywhere, not just as a one-way street towards the GPU.
The move away from eDRAM comes naturally at certain framebuffer sizes. You'll recall they were aiming high (1080p standard), which didn't come to pass in the end. But you can see that if they thought RSX could give them 1080p+2xAA or some such in the majority of games, the eDRAM proposition must have seemed too expensive to pursue at the time. This assumes that RSX specifics (width, clocks) were finalized a while after the decision of no more eDRAM was settled.
Heinrich4 said:
Indeed ,but maybe the x360 is more like the gamecube with gpu faster access to main RAM (PS2 only 1.2GB/sec gs and gamecube is 2.6GB/sec).
Yeah, definitely the closer comparison, due to the same PPC CPU family and a GPU from the same company.
From the GPU onwards, PS3 is quite similar to Xbox 1 architecture actually, with the relatively bandwidth constrained graphics hanging off a UMA. Of course the secondary memory pool and Cell makes it different again.
I am sure Sony wanted eDRAM. But they got NV as partner, and you know NV stance on eDRAM, they're aren't as open to it as ATI. I am pretty sure Sony asked NV to add eDRAM to RSX and I am sure NV persuaded Sony away from eDRAM. Sony invested a lot of money into eDRAM before PS3.
I still think Sony should have just used something like Xenos daughter die in functionality for their GPU and clock the Broadband Engine at 4 GHz and stick with all eight SPUs despite bad yield. I think it'll make cheaper PS3, keep BC, and create their end game Playstation architecture. Sticking RSX in PS3 is just messy, they’re at the mercy of NV in regard to BC. If they don’t go with NV, I think PS4 won’t be compatible with PS3.
I am sure Sony wanted eDRAM. But they got NV as partner, and you know NV stance on eDRAM, they're aren't as open to it as ATI. I am pretty sure Sony asked NV to add eDRAM to RSX and I am sure NV persuaded Sony away from eDRAM. Sony invested a lot of money into eDRAM before PS3.
I still think Sony should have just used something like Xenos daughter die in functionality for their GPU and clock the Broadband Engine at 4 GHz and stick with all eight SPUs despite bad yield. I think it'll make cheaper PS3, keep BC, and create their end game Playstation architecture. Sticking RSX in PS3 is just messy, they’re at the mercy of NV in regard to BC. If they don’t go with NV, I think PS4 won’t be compatible with PS3.
How would that make it cheaper?
Also,i dont think they went with eDRAM from day one,they probably had their own gpu in mind,it did not gave performance they wanted so they went to Nvidia and bought best possible of shelf gpu.
By not having to inclue EE/GS + RAM for one. Which make smaller mobo and thus smaller PS3. And not having to pay NV for RSX would go a long way too. Also PS3 would come with only 256MB of RAM would reduce cost some more. Beside losing BC is a huge blow, bigger blow than what Sony will admit.
They already have GS32 made anyway all they have to do was shrunk it from 0.18u to .09u and bump the clock speed. Who knows maybe they can double the GS32 to GS64 and added some necessary features to support Cell in the process. But this way it would be less powerful than the current PS3, but at least it'll be cheaper at launch and the long run, as of now Sony is screwed, they lost large portions of their market share to Nintendo and MS and been losing money for several years on PS3.
Also,i dont think they went with eDRAM from day one,they probably had their own gpu in mind,it did not gave performance they wanted so they went to Nvidia and bought best possible of shelf gpu.
They went to NV with set of requirements, I'm sure BC is on the list too. NV gave them an old revision of their GPU was not something Sony had in mind. I'm sure when Sony discovered RSX can't do BC all hell broke lose with their NV partnership. Then the broken scaler things must be the nail in the coffin. But for the future Sony's cards are limited. Going all Cell architecture is dead because the chance they won't get support is very high. They had that one chance to do it and they blew it.
On hindsight, Sony should have just released an overclocked PS2 with more RAM for HD support (which they already trial with GScube), eyetoy, motion controller and Bluray and call it PS2 HD or something. And halt Cell for when it's ready.
I am sure Sony wanted eDRAM. But they got NV as partner, and you know NV stance on eDRAM, they're aren't as open to it as ATI. I am pretty sure Sony asked NV to add eDRAM to RSX and I am sure NV persuaded Sony away from eDRAM. Sony invested a lot of money into eDRAM before PS3.
I still think Sony should have just used something like Xenos daughter die in functionality for their GPU and clock the Broadband Engine at 4 GHz and stick with all eight SPUs despite bad yield. I think it'll make cheaper PS3, keep BC, and create their end game Playstation architecture. Sticking RSX in PS3 is just messy, they’re at the mercy of NV in regard to BC. If they don’t go with NV, I think PS4 won’t be compatible with PS3.
I read something about this and like patent september /2002 and if not my mistake with gpu cell like (Visualizer) with PPUs + SPUs and pixel engines with 32 or 64MB eDRAM.
Developers will not sleeping many days but today maybe this ps3 gpu cell like with some nvidia* help(TMUs,compression techs,ROPs etc for pixel engines) could be a monster with modern tecniques created for SPUs(Spurs,phyreengine..).
*(Jen Hsu Huang talk a lot about cell since may 2003...and US$100 million sony pays to nvidia is to high for RSX...)
Sony had no idea that a barely upgraded system like the Wii was going to take off like it did. The Wii was a perfect storm of sorts.
I've always been personally fascinated with the idea of the EE having 2 more or so VUs as well as more RAM to cope with the loads. It would've certainly given the PS2 an edge in managing complex scenes with alot of physics and polygons. Had the PS2 been equiped with a "real" GPU with it's own dedicated T&L and left the EE to do all the non graphics work, I can only imagine how realistic and interactive worlds could've been at the time of the PS2's release. However, had that happened, the Xbox surely would've been equiped with something faster than a paltry 733 MHz P3 derivative, maybe a P4 1.6 GHz assuming MS could've afforded to lose even more just to manufacture the console. I always find the EE vs X-CPU debate interesting. Poor EE got stuck aiding the GS.
Sony should have that idea when they looked at their Eyetoy. Eyetoy was like one of the best selling peripheral. They just didn't know how to move forward with it. Sony Move controller tech was demonstrated long time ago in one of the dev conference. Eyetoy ability to track colour bulb thing look promising even back then for gaming. They sit on that idea and didn't do anything with it. Wii happened and Move is still not out. That's how clueless they were.
I've always been personally fascinated with the idea of the EE having 2 more or so VUs as well as more RAM to cope with the loads. It would've certainly given the PS2 an edge in managing complex scenes with alot of physics and polygons. Had the PS2 been equiped with a "real" GPU with it's own dedicated T&L and left the EE to do all the non graphics work, I can only imagine how realistic and interactive worlds could've been at the time of the PS2's release. However, had that happened, the Xbox surely would've been equiped with something faster than a paltry 733 MHz P3 derivative, maybe a P4 1.6 GHz assuming MS could've afforded to lose even more just to manufacture the console.
I don't think more VUs would work. The EE is not arranged neatly so I don't think it can scale that well. The EE was quite capable at doing T&L though. Beside the timeline in PS2 development, is that Sony developed the GS first. Then they designed something that could feed it, that's the EE. They had to partner Toshiba to do it. If GS had T&L, the EE wouldn't be as powerful. Also I think PS1 was the same, they developed their GPU first and choose a CPU for it. Furthermore Sony still experimented with GPU in their GSCube, EE was the same but the GS got bigger eDRAM. So making fast raster hardware is within their capability and was still trying.
That change for PS3, where they pour much of the development effort into Cell and tacked RSX at the end. My guess is that Sony had a raster hardware for GPU ready to pair with Cell for quite sometimes. Why they went with RSX is said to be politic, as in devs pressure for familiar dev environment. Anyway RSX was definitely last minute, you can see it when NV didn't have RSX ready when it was announced and devkit had the 256 bit version and of course later they have to add the PS2 hardware in there too, subsequent delay, second revision still having GS for BC, and BC canned at the end.
I always find the EE vs X-CPU debate interesting. Poor EE got stuck aiding the GS.
An interesting debate would be Cell Vs Xenos parent die in term of shading prowress. They are similar in transistors count and FLOPS and it would be a nice comparison for how to go forward from here.
For one I still haven't bought into GPU will process everything in the future. To me writing code for other things than graphic in GPU is even more of a pain than say Cell. Sure you can write simple matrix stuff on it, but when it need to be real apps its a lot harder and most likely cost more to develop, than just writing it for CPU or Cell even. At the moment GPU taking over is NV things, even AMD still pair their Fusion with CPU cores.
I see x86 + GPU fusion like amalgamations for the future. The more general purpose capability the better, as it's always needed with as many tasks as a modern CPU is handling. With either a built in GPU or similar hardware, outboard GPUs won't be necessary anymore except for heavy tasks that really require it like gaming, 3D rendering, GPGPU farming, etc. This could should bring down the cost of motherboards, except for the initial intro of such chips, as a huge leap in memory bandwidth will be needed to get the full potential out of a Fusion like part. It really seems like processors, OSs, the programs that operate in/on them, can never really be thrown away only for a whole new system to be developed, at least in the consumer computer space. Things have to be added on, while retaining some legacy compatibility until the legacy silicon is deemed unnecessary, and that which was exotic becomes the new legacy hardware. Fusion like hardware could also take a huge chunk out of dedicated graphics boards in terms of sales, however I think there will be enough gamers out there who would want dedicated boards, as Fusion types/hybrids could go off on all sorts of tangents in terms of "x" amount of "general capability" paired with "x" amount of "GPU/"accelerated"/whatever capability", and honestly it has the potential to be a consumer nightmare for the uninformed, though the general consumer already is
Sony should have that idea when they looked at their Eyetoy. Eyetoy was like one of the best selling peripheral. They just didn't know how to move forward with it. Sony Move controller tech was demonstrated long time ago in one of the dev conference. Eyetoy ability to track colour bulb thing look promising even back then for gaming. They sit on that idea and didn't do anything with it. Wii happened and Move is still not out. That's how clueless they were.
at the Engadget Show some months ago, Richard Marks said, most of the Eyetoy options were never used, because they needed too much CPU power to make them work. For example the headtracking used something like 30% of the EE to work. No wonder we didn't see too many "mature" games come out of this. Going forward though, they should've pushed the PSEye MUCH more than they did in the past 3 years. I mean... ONE retail game using it (beside some gimmicky usage in games like Burnout Paradise) and only some mediocre, lower priced PSN titles.
No wonder third parties didn't make any games for it, since ... if Sony doesn't push it, why should they, on a platform that seriously underperformed in the beginning. Move will change this to a degree, since many people will own the camera now, thus "Kung Fu Live" or whatever it is called comes to PSN too...
How's so? The only commonality is the eDram. Apart from that the 360 is something between the original XBOX and the PS3 (unified memory, a GPU very close to PC GPU and a processor that uses in part some of the technology found in Cell).
The PS3 reminds me more of the PS2 because the Vector Units are replaced by the SPU's and the CPU is used to assist the GPU.
PS3 is as unlike PS2 as XB360 is. PS2 didn't even have a proper GPU, with half the workload designed to be performed on EE - in essence a two-chip GPU solution. RSX is a standalone GPU but is helped to achieve more with the SPEs, which can be optionally purposed for any job unlike PS2's VU's which, if they never processed vertices, would result in nothing getting drawn!
