Quad Core RSX GPU!?!

[Nerve-Damage]

This part of thread is getting interesting…………… (http://www.beyond3d.com/forum/showthread.php?t=26620&page=5) :spiny:

Credit: Version
http://web.axelero.hu/varga1973/rsx.JPG

IIRC it's a picture from some patent by Sony or Toshiba
Version..you're my personal hero

I wonder will Sony incorporate some of the technology from the Sony/Toshiba design into the Nvidia/Sony RSX design!?!

[CrumCon]

Does it works like PC GPU in SLI ?

But like many people said in many threads, the possibility of spec changes are very minimum.

[Nerve-Damage]

Does it works like PC GPU in SLI ?

Pretty much; except everything is on one GPU die. This type of design is no different than say a multicore CPU design. This type of designs tends too help reduce latency among multiple processing units compared to having them spread across the system in either separate sockets or slots.

But like many people said in many threads, the possibility of spec changes are very minimum.

Spec change?

The truth is we don’t know the specs per-se about the RSX setup; meaning how many pipelines or the possible amount of graphic cores that are sitting on the GPU die. We're just going off assumptions of what we know about the G70 GPU family designs.

Sony has never mention or stated that the RSX is a single core design. So there’s still a possibility for an multicore RSX GPU design (i.e. Daisy Chain).

[rpgamer_2k5]

^^^^^ We weren't told that the RSX is a G70 derivative. The G70 has been out for long, 3 variants are out and yet we have not heard that the RSX is another variant. Besides the G70 was a light project since Dr. David Kirk was not overlooking the project.

I can't wait til February...

[Gegenki]

But how do we know thats the RSX
Looks to me like a general purpose multicore processor
I'd have thought the RSX if multicore would be very similar to the cell, to aid in its ability to integrate and to promote the cell technology

[Saibo]

I thought creating a multi-core GPU wasnt possible yet? Otherwise i would guess multi-core configuration long time ago(it would only compliment the Cell design). IMO i dont think the RSX is just a G70 with some tweaks, and features taken out. I think Sony and Nvidia started from scratch.

[Nerve-Damage]

I thought creating a multi-core GPU wasnt possible yet? Otherwise i would guess multi-core configuration long time ago(it would only compliment the Cell design). IMO i dont think the RSX is just a G70 with some tweaks, and features taken out. I think Sony and Nvidia started from scratch.

Well multi-GPU (Multiple GPUs on a single PCB) boards have been out since the 3DFX Voodoo days. It’s just recently that rumors are starting too pop-up of multicore GPU designs. ATI more specifically is rumored to have a multicore GPU out next year (2006) and there is some rumbling from the Nvidia camp also about multicore GPU designs.

[xbdestroya]

I don't think it's that creating a multi-core GPU isn't possible, just more a matter that in a way all GPU's are multi-core (or multi-quad, etc...) in a sense. I mean at the same die area, what do multiple cores on a GPU do for you that a single larger GPU core wouldn't be able to do?

By the way, I think that drawing is from the 'Visualizer' patent from back in the day, not sure though.

[Domination]

I still have mixed thoughts on this.

[Nerve-Damage]

Here’s a picture of my two (SLI) 3DFX old schoolers with multiple GPUs on the PCB. Man I came along way from running those too now running (SLI) eVGA 7800GTX monsters. :drunk:
http://img51.imageshack.us/img51/6839/im0002194rg.jpg

[CrumCon]

PS2 hardware doesnt end well up with many devs, which is design by both Sony and Toshiba, i hope they dont make the same mistake like they did with ps2, cause its so hard for many devs to work with ps2 hardware

[Revenge]

-Global Cache -> Turbocache, through XDR Memory.

-Local Cache -> Cache L1 and L2

-Memory Module -> GDDR-3 memory module controled directly by the GPU

Gaif ->> Compusite Engine like the Gforces 6xxx and 7xxx that suport SLI.

Why? Because this would be to compose rendered images for more than 1 GPU's into only one (in this case its 4), what it implies a GPU Multi-Core.


However, as the actual GPU's can already be considered multi-Core (independent pixel pipelines and vertex pipelines) the only reason I see with some logic it would be if the GPU had 4 dies, this means, 4 separated cores but in the same package.
U can see the diference between the Pentium D and the new Presler, one have 2 cores in the same die, the other one have 2 dies.

This being true, the RSX will be monstrous, and the only way to keep it economic was dividing it in 4 parts and use (quad)SLI so they can work like one.

In certain way, eDRAM of Xbox 360 was also manufactured with this intention, even so its not a co-processor in capacity parity, as RSX.

Other interesting speculation is the fact of memory GDDR-3 (that is obviously slower than the XDR) to be used in part to have "free Antialiasing" as in Xbox 360, blocking for example one of the modules of memory GDDR-3 for effect, and using XDR as front buffer(as the local memoryl in the 6200 Turbocache, that always lodges what it sees in the screen, cause its faster than the memory RAM of the system, which is used for cache, geometry, textures and other things).

It seems to be more flexible than simply conect GPU's in parallel, and more cheap than eDRAM (its too expensive) separately to obtain AA without have loss of performance, being able to even break the limit of x2/x4 AA of Xbox 360.

With all this we can verify that the RSX don't have anything to do with the G70/G71 and that the RSX works much more with the CELl and the XDR memory than we thought.


PS: Im Portuguese, so sorry for my bad english.

[xbdestroya]

Well, I still thinkit's the Visualizer patent, not anything RSX-related per se. I would take this as 'proof' of anythign just yet. BUT, Revenge where'd you sneak in from? Good post (even though I disagree), welcome to the forum. :smoke:

Exactly, GPU's are already multi-core in a sense, the reason to go multi-die would be for die-size/yield reasons.

Also Nerve, though I'm not surprised it might have flown under everyone's radar, Version actually posted that drawing yesterday in the current semiconductor thread, post #9. :)

http://forums.e-mpire.com/showthread.php?t=49646

[Saibo]

Well, I still thinkit's the Visualizer patent, not anything RSX-related per se. I would take this as 'proof' of anythign just yet. BUT, Revenge where'd you sneak in from? Good post (even though I disagree), welcome to the forum. :smoke:

Exactly, GPU's are already multi-core in a sense, the reason to go multi-die would be for die-size/yield reasons.

Also Nerve, though I'm not surprised it might have flown under everyone's radar, Version actually posted that drawing yesterday in the current semiconductor thread, post #9. :)

http://forums.e-mpire.com/showthread.php?t=49646

I know it was part of the Visualizer patent, but Sony is still apart of the RSX gpu design, So maybe both Nvidia and Sony can still put there minds together to create a new Visualizer? It'll still be a Visualizer wont it? Kinda of like how STI create Cell, orginially Toshiba only wanted the SPE, but IBM insist on having a PPE.

Silly/stupid question, going mutl-core would be cheaper, but would it be still cheap it each one of the core were design with specific features in mind, and being dedicated to that on feature, something like displacement/relief mapping on one core, second core for something else(lighting?), third for something else, 4th core a custom G70 gpu? :eek2: or is that just BAD multi-core GPU design idea on my part?

If you look at games like KillZone and MotorStorm they just scream displacement, relief mapping on many areas. Im a huge fan of micro poly displacement mapping, if the PS3 could do it with no sweat, it would make my year! :D

[Nerve-Damage]

Also Nerve, though I'm not surprised it might have flown under everyone's radar, Version actually posted that drawing yesterday in the current semiconductor thread, post #9. :)

http://forums.e-mpire.com/showthread.php?t=49646

LOL...... thanks for the point-out xbdestroya

[xbdestroya]

I mean we'll still keep this thread going though, since it's more focused on that aspect of things. :)

@Saibo: I'm not going to lie Saibo, I really don't know. :smoke:

All of that is stuff that really is what GPU's are about anyway, and I would say that in a sense, vertex shaders kind of represent the 'specialized' hardware on GPU's that have brought us to harsware T&L. I see what you're saying though, if some sort of multi-core on-die arrangement with seperate specialized cores would make sense or not as an alternative is a question that would have to be fielded by someone else. Maybe Cpi can approach it more definitively.

[VG Aficionado]

Here’s a picture of my two (SLI) 3DFX old schoolers with multiple GPUs on the PCB. Man I came along way from running those too now running (SLI) eVGA 7800GTX monsters. :drunk:Hey Nerve, those are Voodoo 2 cards, aren't they? 12MB models, huh? I bought and installed one of those in my old Pentium II PC and games performance was CPU limited. Later on I wanted to add another card to make an SLI system, but I never found the very same card. Doesn't really matter since performance wouldn't have been much higher, although I could have achieved 1024x768 resolution instead of 800x600 max.

Well multi-GPU (Multiple GPUs on a single PCB) boards have been out since the 3DFX Voodoo days. It’s just recently that rumors are starting too pop-up of multicore GPU designs.I think one of those processors is a geometry unit and the other two are texturing units. The three of them form a single GPU. Or that's how I believe it works.

[GTShotoKen]

Does anyone think that the RSX might have features derived from the G80, even though it's a card designed for Shader Model 4.0?

Or at least support all of the features of Shader Model 3.0?

Most SM 3.0 compliant cards today don't support all of its features right?

This is all dependent on if any of these features are usable through openGL though.

[Nameless]

I made this statement some time ago...
It would not surprise me if the RSX is a hybrid of the cell, resulting in a multi-core GPU. Nvidia could be waiting to the last moment to release this info to keep a competitive advantage from an IP perspective. You heard it here first! Peace

[Nerve-Damage]

Hey Nerve, those are Voodoo 2 cards, aren't they? 12MB models, huh? I bought and installed one of those in my old Pentium II PC and games performance was CPU limited. Later on I wanted to add another card to make an SLI system, but I never found the very same card. Doesn't really matter since performance wouldn't have been much higher, although I could have achieved 1024x768 resolution instead of 800x600 max.

Yupp, Voodoo 2(s) with 12MB of EDO Ram :rockon: :eek2:


I think one of those processors is a geometry unit and the other two are texturing units. The three of them form a single GPU. Or that's how I believe it works.

Well yeah you’re pretty much right!! However 3DFX would always flaunt them as independent units or processors. However I always felt that ATI was trully the first to created a multi-GPU based graphic card (Rage Fury Maxx cards).

[xbdestroya]

Does anyone think that the RSX might have features derived from the G80, even though it's a card designed for Shader Model 4.0?

Or at least support all of the features of Shader Model 3.0?

Most SM 3.0 compliant cards today don't support all of its features right?

This is all dependent on if any of these features are usable through openGL though.

There's some thinking that RSX might have some share of G80's feature-set included, but unfortunately no one has anything to go on that's solid other than the time it's taken this chip to tape out.

[cpiasminc]

Back before E3 and well before G70 was announced, there was that conference call with nVidia, where Jen-Hsun Huang said that the GPU for PS3 would be (to summarize)"a one-off design based on their next generation product line." I think it would have been fair to assume at the time that he was referring to the immediate successor to the then top-of-the-line 6xxx series.

More than anything else, I really expect just a raw power difference. The CPU-GPU bandwidth alone (as opposed to PCI-E) can have some interesting impact in the effective throughput.

[rpgamer_2k5]

I made this statement some time ago...
It would not surprise me if the RSX is a hybrid of the cell, resulting in a multi-core GPU. Nvidia could be waiting to the last moment to release this info to keep a competitive advantage from an IP perspective. You heard it here first! Peace The Cell-RSX hybrid solution has been on this forum for quite some while. The SPEs would make excellent vertex processors for example; and would be excellent if mated with a NV42 pixel pipe.

We'll have to wait and see, though.

[CrumCon]

during E3 conference, that Nvidia CEO said RSX will be based on their next-gen GPU.. so i doubt RSX is CELL hybird. Unless Nvidia will pay license to use Cell technology for their next-gen GPU.

Nvidia will use their own tech, license tech from outside is not their style.

[xbdestroya]

during E3 conference, that Nvidia CEO said RSX will be based on their next-gen GPU.. so i doubt RSX is CELL hybird. Unless Nvidia will pay license to use Cell technology for their next-gen GPU.

Nvidia will use their own tech, license tech from outside is not their style.

Well it's not really like *that* - this is going to be Sony's GPU afterall, not NVidia's. It'll be Sony doing the licensing, and in theory if they worked out some cool tweaks to NVidia's own architecture, they could surely use it no problem. As for NVidia, their future GPU's won't really have any of the tech developed specifically for RSX come back at them, so there won't be a licensing issue there.

[CrumCon]

There will be licensing issue if their next-gen PC GPU based on Cell hybird.

if RSX is indeed cell hybird, could it be possible that both RSX and Cell running in SLI together? both has the same amount of memory.. !?

[Nerve-Damage]

during E3 conference, that Nvidia CEO said RSX will be based on their next-gen GPU.. so i doubt RSX is CELL hybird. Unless Nvidia will pay license to use Cell technology for their next-gen GPU.

Nvidia will use their own tech, license tech from outside is not their style.


The funny thing about using certain words, especially in a manipulated fashion; is that they can mean something totally different from person to person. If Nvidia had stated that the RSX is just a G70 or G80 derivative, and then RSX GPU mystery is pretty much solved, i.e. just an updated Nvidia based chip.

However this isn’t the case with the RSX!?! Nvidia will provide graphic technology based off there latest graphic chip family to Sony. What Sony chooses to do with these chips is up too them; meaning Sony can fashion these chips into GPU die design that may not be so common on a PC graphic card. Like a multi-core GPU if the technology permits Sony to achieve that goal. Augmenting chips isn’t nothing new, it’s the end-user (Sony) application or hardware (GPU) final design that may (or may not) limit the actual Nvidia chip performance.

From what I understand, Nvidia will provide the heart & soul (Graphic Cores), while Sony will provide the vessel or body (GPU Die and chip arrangement on the die).

[Infernal]

Guys guys I think everyones forgetting something. If the RSX really is more powerful than were led to believe wouldnt they put 2 7800GTX cards in SLI in the dev-kits and not just a single card? (For those that dont know the dev-kits originally had 2 6800GT's in SLI and now its just a single 7800GTX). The dev-kits are meant to represent the final power as best as possible arent they?

[VG Aficionado]

Guys guys I think everyones forgetting something. If the RSX really is more powerful than were led to believe wouldnt they put 2 7800GTX cards in SLI in the dev-kits and not just a single card? (For those that dont know the dev-kits originally had 2 6800GT's in SLI and now its just a single 7800GTX). The dev-kits are meant to represent the final power as best as possible arent they?You are talking about alpha and beta kits respectively. Since both kits have a 2.4 GHz Cell processor, I don't consider they're representative of PS3's full power. Final dev kit will feature 3.2 GHz Cell, RSX, Blu-ray and whatnot, and only when we see the first games running real time off these dev kits or on the final PS3 unit, we'll know how they will really look like.

[Nerve-Damage]

Guys guys I think everyones forgetting something. If the RSX really is more powerful than were led to believe wouldnt they put 2 7800GTX cards in SLI in the dev-kits and not just a single card? (For those that dont know the dev-kits originally had 2 6800GT's in SLI and now its just a single 7800GTX). The dev-kits are meant to represent the final power as best as possible arent they?

Actually there is an actual statement (running around the net) from Mark Rein hinting that (EPIC) was using a SLI setup of the newest Nvidia graphic technology in the PS3 Alpha Kits during the E3 PS3 UT2007 demo presentation.

[xbdestroya]

There will be licensing issue if their next-gen PC GPU based on Cell hybird.

if RSX is indeed cell hybird, could it be possible that both RSX and Cell running in SLI together? both has the same amount of memory.. !?

CrumCon I'm not going to spend too much more time trying to explain it so... no, under no circumstance would NVidia pay Sony for licensing. :smoke:

Sony will pay NVidia roughly $5 per RSX chip fabbed for licensing, but not the other way around.

Even if RSX incorporates some aspects of Cell or it's patent or R&D, there is no way that NVidia's next PC chip will, and since Sony will be the owner of the RSX design, NVidia will not have access to it's independent use or sale.

[Nerve-Damage]

Sony will pay NVidia roughly $5 per RSX chip fabbed for licensing, but not the other way around.

Even if RSX incorporates some aspects of Cell or it's patent or R&D, there is no way that NVidia's next PC chip will, and since Sony will be the owner of the RSX design, NVidia will not have access to it's independent use or sale.


100% correct!!

Sony does not have the same deal or contractual agreement with NVIDIA as Nvidia had with Microsoft during the Xbox days. Sony will be producing the bulk (if not all) the graphic cores within its own fabrication facilities. And as xbdestroya pointed out Sony will pay Nvidia a per-chip royalty fee.

What Sony chooses to do with the graphic cores is Sony business...as long as it doesn’t sale/give out any patent Nvidia IP, to Nvidia competitors or infringe on anything that deals with Nvidia business model (Computer hardware mostly). Sony has the right to augment the chips too their specifications or business model. Sony (hypothetically) may have the technological engineers to actually fabricate the Nvidia graphic cores into a dual core, tri-core, or quad core GPU design (or even a hybrid GPU/CPU CELL like design).

Anyhow January is almost upon us………….. :queer:

[cybergrue]

Guys guys I think everyones forgetting something. If the RSX really is more powerful than were led to believe wouldnt they put 2 7800GTX cards in SLI in the dev-kits and not just a single card? (For those that dont know the dev-kits originally had 2 6800GT's in SLI and now its just a single 7800GTX). The dev-kits are meant to represent the final power as best as possible arent they?
It may just be a stop-gap measure to get more dev kits out the door, and not indicitive of the final product.

The other thing to consider is that the Ps3 will support two video out channels (ie, the ability to use two tvs/monitors showing different things at the same time), supposably both running at 1080p. Wouldn't this ability imply multiple cores for the GPU? One, or more GPUs per video channel out.

[Nodieza]

^ That's what logic would say but Sony is the master or wizadry and they can always find an insane way to solve a problem so who knows.

[xbdestroya]

In fact most modern high-end video cards support dual-monitors, so really the two screens at 1080p doesn't indicate multiple chips in and of itself.

[RzrWire]

I'm just gonna post the same thing here as I did in the PS3 Memory Configuration: Get Your CES Bets In!! (http://forums.e-mpire.com/showthread.php?t=49672&page=2) thread just in case it gets missed or overlooked.


Original Post:

I chose the "256MB XDR & 512MB GDDR3" option because if you think about Nerve-Damage's Multi-Core RSX Post it makes sense. Although I vote for a dual core RSX as the most likely option. If one RSX is more powerful than two 7800GTX cards in SLI why not have a dual core RSX running in SLI mode? Now imagine if each core had 256mb of GDD3 memory giving the RSX chip as a whole 512mb to play with. Technically Sony could then get away with saying at E3 2005 that indeed one RSX will have 256mb of GDD3 memory, but failing to mention that it is a dual core design each with it's own separate memory. Now wouldn't this be a welcome surprise!!! :buldge:

[Nerve-Damage]

My Dream!! :eek2:

http://img33.imageshack.us/img33/2817/nervedamagersxgpudream2ky.jpg

[jako]

[QUOTE=Nerve-Damage]My Dream!! :eek2: /QUOTE]

Why not ? at E3 sony's press conference Nvidia said that RSX "will take full advantage of Cell's power", so why not :
RSX="kind of" G70 + "kind of" Cell

[RzrWire]

To be honest I would be more surprised if the RSX wasn't a hybrid of the cell architecture than if it actually is. Only the RSX would use it's cell design specifically for GPU heavy algorithms & tasks.

[rpgamer_2k5]

Blah. Get rid of the N42 vertex pipes and just use the 4 SPEs. One SPE can rival the GTX in vertex operations. Four SPEs would allow for 3.2 billion vertices/sec. I would take that over the GTX-512's 1.1 billion vertices/sec any day. :)

[Nerve-Damage]

Blah. Get rid of the N42 vertex pipes and just use the 4 SPEs. One SPE can rival the GTX in vertex operations. Four SPEs would allow for 3.2 billion vertices/sec. I would take that over the GTX-512's 1.1 billion vertices/sec any day. :)

I was thinking about that; dropping all the VS pipelines out of the Nvidia G70/G80 graphic core, thus increasing the PS pipeline count. However I figured any PS3 games (Post-E3) may actually be using the Nvidia based VS-pipelines (than the SPE) and also figured some developers may prefer to use the Nvidia VS variants until they get up to speed on how to use the SPE(s) more efficiently. So I kept 4 out of the 8 VS pipelines and increased the PS pipelines quite a bit.

[Pistolero]

You are setting your expectations too high. I think of a GPU based on the G70 with functions borrowed from the G80 and customized to get the best out of the Cell.

[Nerve-Damage]

You are setting your expectations too high. I think of a GPU based on the G70 with functions borrowed from the G80 and customized to get the best out of the Cell.


Not really!!

I have no expectations on the final PS3 design or hardware…as long as it works.

What I do have is theories or seeing my dream setup…but no high expectations of setting myself up for a let down.

I’ve been through too many console launches for that too ever happened. ;)

[rpgamer_2k5]

You are setting your expectations too high. I think of a GPU based on the G70 with functions borrowed from the G80 and customized to get the best out of the Cell. The problem is the G70 is a huge GPU. Trying to fit in a G70-512 plus into the PS3 might not even be possible at all. If Sony is smart they would rather do what they usually do; develop a GPU from grounds up for the console. You really don't need the heavy vertex power of the G70-512 for the PS3. That would just be a waste, it would be more efficient to throw in more pixel pipes or a quad.

Keep in mind that Ken Kutaraji did state that the RSX is not based on a PC GPU and that would be a much smarter choice.

That does not mean that they would include SPEs, since 7 on the Cell is more than enough. Physics would not require even 3 SPEs, because we're not going to be see games with complexity of demos being proposed in the academia. If three SPEs of the Cell are tasked for graphics that result would be more than enough. Heck a few would task four for graphics, 2 for physics and one for "AI" but allowing PPE to do most of the AI while 1 SPe supports it would be better. The RSX with its super-pixel pipes (speculation) would run the rest of the show.

Remember the CEO of Nvidia sees the Cell as the real challenge in the near-future rather than the RSX. David Kirk also praises the SPEs.

Nerve-Damage, ever since the introduction of higher level programming on vertex pipes the aim was to turn vertex pipes into general-purpose processors. The SPEs is closest to the ideal in the gaming realm. Even though the SPEs are not as flexible as conventional CPUs currently, they would roll over any vertex processor. :)