Dual Gpu console revisited

[Guden Oden]

The point is TWO 512's would offer a step beyond anything else currently achievable.

And you're not the first to put forth that "point", and not the first who apparantly does not understand the issues regarding such an undertaking. Just for starters imagine the board area that would be taken up by a dual GPU setup. Have you seen pics of ASUS and Gigabyte's twin 6600GT vidcards? They're absolutely enormous. That's just the space needed for GPUs, memory and power regulation. Add to that, Cell itself, its memory and power regulation, southbridge and auxiliary chips (of which there'll be a bunch in the first PS3 revision no doubt, seeing Sony's past hardware designs), and so on. Plus the XIO-to-XIO bridge needed to connect two GPUs rather than just one I might add (serial links aren't PCI buses you can just arbitrarily chain up devices onto).

There's simply no way you're going to be able to fit all that into the current PS3 case. And no, before you think the thought, you can't put them on alternate sides of the mobo either because there isn't space in the board layers for all the extra traces to do that. Besides, on the reverse side of the mobo there'll be the BDROM, the harddrive bay and possibly the internal power supply. No room for such ideas.

I know SLI in a PC uses redundant RAM. This COULD be a big enough issue to kill the idea. However, isn't there a way to virtualize two GPU's as one?

You could say a 7800GTX is a "virtualized" SLI 6800 vanilla. That isn't SLI in any way shape or form though, it's just expanding the chip to include twice as many rendering pipelines, with the drawbacks that includes with regards to costs and such, manufacturing really huge silicon dies. So short answer to your question's going to be "no, not really". Not with an immediate-mode renderer anyway. Had Nvidia been making a tile-based deferred renderer for PS3, two or more GPUs might have been able to share memory at least somewhat efficiently. An immediate mode renderer accesses memory in a very chaotic manner (one of the reasons MS didn't go with a fully UMA design this time 'round), with two sharing the same RAM they'd end up stepping on each other's feet constantly, not to mention likely starving for bandwidth even assuming memory contention wasn't an issue.

SLI is just a term I'm using. It could be implemented differently. SLI has limitations based on the PC.

It's not limitations based on the PC per se. It's limitations inherent with rendering different pixels on different GPUs into different memory. It'd appear on any other platform other than PCs that was using SLI.

In the case of render to texture effects, there's no way to guarantee that the pixels you'll need on GPU A will actually have been rendered by GPU A. They might sit in GPU B's memory and do you absolutely zero good.

There are ways around that as I implied in my original reply, you can duplicate all render to texture work on both GPUs for example, but that would destroy the efficiency you claimed would magically appear from doing SLI in a console.

And Guden Oden, funny, I just read Nvidia CEO claiming "SLI will become more important going forward", not less.

Your reply shows you simply did not understand the point I was making, as the non sequitur you present has no bearing on any of what I said.

You might also want to stop and think for yourself for a second there before you blindly accept the word of the CEO of a company that has a vested interest in seeing SLI become prevalent. That's not saying the tech doesn't have a future, but rather just training yourself in recognizing propaganda when you see it.

I agree SLI has problems in a PC setup, again that's not really my point.

And my point is you merely theorize that the problems of SLI are bound only to the PC platform (and present no evidence to support that claim), when in actuality they're not.

 

[pc999]

The only console that I think it had a little chance to have dual GPU is Rev (a pricier version of it) to pease HDTV fans (as a add on or out of the box) which the only difference would be the rez, but even then I gave it few chances till it got axed for good.

 

[Shompola]

Bill, you would be better off if you started a thread discussing why SEGA choose to design an arcard board, Lindbergh less(atleast on paper) powerful than the PS3 and Xbox 360.

 

[London Geezer]

Bill, you would be better off if you started a thread discussing why SEGA choose to design an arcard board, Lindbergh less(atleast on paper) powerful than the PS3 and Xbox 360.

He did more or less, and it got deleted (not even locked, DELETED.. that's quite something.. but it wasn't Bill's fault) after a major highjacking.

 

[Megadrive1988]

I was hoping that a new high-end console would be made, like NEO-GEO but for modern 3D. using something like 4 highend DX10 / SM4.0 GPUs in parallal. it would blow away the PS3 much like the NEO-GEO blew away the cheaper consoles of its generation.

$1000 console with $50 games that showed a large difference over X360/PS3. and one where you have a variety of input devices including V.R.

edit: oh yeah, and in a console environment, 2 or 4 GPUs would not use SLI as we see in PCs - in a console, multipul GPUs would be arranged in parallal and would communicate with each other in a much more efficient, tighter mannor, much like the in arcade boards where several GPUs are used.

 

[wireframe]

I was hoping that a new high-end console would be made, like NEO-GEO but for modern 3D. using something like 4 highend DX10 / SM4.0 GPUs in parallal. it would blow away the PS3 much like the NEO-GEO blew away the cheaper consoles of its generation.

$1000 console with $50 games that showed a large difference over X360/PS3. and one where you have a variety of input devices including V.R.

I wouldn't mind seeing something like this myself, but I think it is clear the market as a whole would not. Therefore, this console would cost $3,000+ and because it is so powerful it would demand games that are even more complex than what we have today so the development costs would skyrocket. Because the console is so expensive, few buy it so there are few potential customers for each game, so....let's say the games are $1000+ a pop.

Now, I am not saying some of those games may not be absolutely worth $4,000 to play, but this isn't exactly a risk-free proposition for anyone involved.

I hope my example helped.

PS. Just to drive the point home for sure: Neo-Geo games were created at a time when there really wasn't a huge cost delta between games in "real money" and even individuals could create a game. That, today, would still yield Pac-Man

 

[Tim]

This is true.

However it does nothing to negate the fundamental value of the idea.

The main problem with the idea, is that it is a incredible ineffiecient and expensive way to add performance and worse the cost of multichip solutions would scale much worse over the lifetime of the console than a single chip solutions.

To add an extra RSX in a PS3(not that it actually would be possible) , they would not only need an extra chip they would also need another 256MB RAM, more voltage regulators, better power supply, more cooling, larger enclosure and a more complex PCB with wastly more traces. All this for a performance increase that at best would move the graphics performance a year forward.

So now we are able to get 2007 performance in 2006. In 2007 it would still have the graphics performance of a 2007 console but be around 30% more expensive to produce than a 2007 console(and the %-difference cost won't decrease over time, it will more likely increase). So the result is one year of performance advantage and 6 years of production cost premium.

 

[Lysander]

The X360 has driven home for me that consoles have lost that wow factor.

wow, you lost some faith there; do you care to explain?

 

[jvd]

Bill please follow this


Dual gpu in console means


1) Double the gpu costs

2) Double the ram costs

3) Bigger form factor

4) Higher power draw

5) More expensive cooling

6) louder cooling

7) Lower production numbers( u now have to make 2 gpus for every console cutting production rates in half )


So tell me why would a console company make a dual gpu console ?

 

[Megadrive1988]

I don't think a dual-GPU console would be THAT difficult to implement at a reasonable cost.

Saturn had dual graphics processors (VDP1 and VDP2) in addition to dual CPUs and a bunch of other processors. ($400)

the 3DO M2 had dual-CPUs ($300 to $400

the original 3DO had dual 2d graphics chips ($700)

NEO-GEO had a bunch of custom chips ($400 base unit)

NEC SuperGrafx had dual graphics chips ($300 in Japan)

granted, none of these consoles were very successful but it does show examples of consoles with more than one GPU and/or CPU.


I believe that mainstream next-next gen consoles are going to contain two or three 'CGPUs' (compute-graphics-processing-units) which are equally capable of processing a game and rendering its visuals.

 

[Bobbler]

The only way a dual gpu works in a console is it isn't the standard crappy SLI configuration (which has huge cost overheads along with not being efficient at all, really).

If they made a chip with the idea that there would be two of them then a dual GPU configuration would be no worse than a single chip solution. In which case it wouldn't necessarily be better. SLI in consoles is a horrible idea and flies in the face of console design, which is made to be as cost effective as possible (since cost has to be eaten for a while). There have been a few consoles with dual graphics as Megadrive1988 pointed out, and the only reason they were feasible is because they weren't a hackjob SLI job. The only reason SLI even works in the PC realm is because its cheap to implement on a software level and marketable to people who don't care about cost vs performance.

If you are willing to throw cost of production and implementation out the window there are far better ways to get performance that would be unparalled than doing an SLI rig. So, I'm not really sure what the point of this thread is, its a worst case example of a situation that won't happen.

 

[Megadrive1988]

forget the term SLI - be it the 3Dfx version or the Nvidia version. Any dual-GPU console would probably never use SLI, especially if a true 2x performance over single CPU is desired. think NAOMI 2 or any of the many of the non-PC based arcade platforms that used more than one GPU.

 

[Bobbler]

forget the term SLI - be it the 3Dfx version or the Nvidia version. Any dual-GPU console would probably never use SLI, especially if a true 2x performance over single CPU is desired. think NAOMI 2 or any of the many of the non-PC based arcade platforms that used more than one GPU.

I agree, thats part of what I was trying to say (SLI is garbage, and they'd need to make a chip that was planned from the beginning that way). Also, arcade boards are usually sold at a profit and cost several thousand monies, so we can't feasibly compare arcade to console. And if we are comparing a situation like that to a console, there are ways other ways to get more performance that should be looked at as well (more ram, etc). The entire topic doesn't fit into the realm of possibility on a console -- consoles are made to be cost effective and saying Sony or MS should have put two GPUs with the power of RSX/Xenos in their console (assuming it was designed from the beginning for efficiency) is silly -- I don't think two GPUs is out of the question, as its been done before, but those two GPUs together are usually comparable to what a single GPU was capable of (at least within a short while, at least no better than what could have been done with a single core if the resources were put into it instead). The only reason I could see for going dual chips (from a company perspective, like Sony's or MS') would be to cut fab costs, by allowing for increased yields of requiring two smaller chips instead of one larger, and chances are with the packaging costs you probably wouldn't save much (if at all) -- otherwise it just doesn't make any sense and, as I said before (and many others), flies in the face of the design of a console.