Support NeoGAF

[eeyore the donkey]

Can anyone elaborate on this... "7900/7800 have 256-bit memory bus and RSX only has 128, so this news can't be true."

Therefore RSX must be weaker, budget gpu, Geforce 2, etc

 

[BlueTsunami]

Can anyone elaborate on this... "7900/7800 have 256-bit memory bus and RSX only has 128, so this news can't be true."

Therefore RSX must be weaker, budget gpu, Geforce 2, etc

It was explained @ B3D a few days ago. Something along the lines of the RSX DOES have a 256-bit Memory Bus but its split between two memory pools. I'll see if I can find the direct quote.

 

[fronn]

Can anyone elaborate on this... "7900/7800 have 256-bit memory bus and RSX only has 128, so this news can't be true."

Therefore RSX must be weaker, budget gpu, Geforce 2, etc

Not exactly true. It isn't just exclusively a 128bit bus, there's the flexIO as well.

It's safe to best against whatever foolhearty suggestions deadmeat and pals make.

I'd also say "Don't worry."

 

[gofreak]

Can anyone elaborate on this... "7900/7800 have 256-bit memory bus and RSX only has 128, so this news can't be true."

Therefore RSX must be weaker, budget gpu, Geforce 2, etc

Well, what's a 7900 with a 128-bit bus to VRAM? Or 128-bit bus to VRAM and 35GB/s to Cell?

Hint: It's not a Geforce2, or a 7600GT It's just RSX.

I think it's fairly safe to say that RSX is a 8:24 G70 shrunk to 90nm with some tweaks, with FlexIO and 22.4GB/s to VRAM. Why the persistent attempts in drawing a 1:1 mapping with a PC card? It's not really possible, in a perfect sense. I might say it has the VRAM bandwidth of a 7600GT, but the shader power of a 550Mhz 7800/7900. And CPU bandwidth many times either. It's kind of fruitless trying to equate it exactly with anything in PC cards. It bears many similarities to various GPUs there, but none are a perfect match.

 

[Pimpbaa]

Can anyone elaborate on this... "7900/7800 have 256-bit memory bus and RSX only has 128, so this news can't be true."

Therefore RSX must be weaker, budget gpu, Geforce 2, etc

Geforce 2? Budget? A 128 bus doesn't neccessarily mean budget. The Geforce 7600 GT has a 128 bit bus for example. The RSX can be extremely powerful at shader calculations since the bus won't limit that much. The only time the 128 bit bus might be a problem is under fillrate intensive situations (like say anti aliasing).

 

[BlueTsunami]

There is a flawed assumption here that the RSX only has a 128-bit bus. It actually has two of them - Sony didn't so much halve the bandwidth as split it between two paths (one to the vram, one back to Cell).

It's not the same as having a 256-bit bus, but it's not nearly as bad as only having a single 128-bit one. With good planning and management of resources, more bandwidth can be had. It also means that the system has pretty good bandwidth for keeping the GPU fed by the SPEs, something that might not be possible if the system had stuck with a slower path between GPU and CPU just to get more GPU->vram bandwidth.

I don't like to pull quotes from other boards but I thought this would add to the conversation.

 

[eeyore the donkey]

It's funny how in 5 minutes people can debunk what the Xbox crew at OA has been preaching for the better part of a year. :lol

Thanks!

 

[BlueTsunami]

It's funny how in 5 minutes people can debunk what the Xbox crew at OA has been preaching for the better part of a year. :lol

Thanks!

Thats why I stopped posting over there, my brain seriously started to hurt. Its like going from one board where actual developers are present and then another board where FUD is rampant. It was tainting my perception on technical aspect of things.

 

[RaijinFY]

Thats why I stopped posting over there, my brain seriously started to hurt. Its like going from one board where actual developers are present and then another board where FUD is rampant. It was tainting my perception on technical aspect of things.

Wow I admire you started, actually. On my side, I even couldn't do it...

 

[Razoric]

It's funny how in 5 minutes people can debunk what the Xbox crew at OA has been preaching for the better part of a year. :lol

Thanks!

:lol :lol

 

[mckmas8808]

There is a flawed assumption here that the RSX only has a 128-bit bus. It actually has two of them - Sony didn't so much halve the bandwidth as split it between two paths (one to the vram, one back to Cell).

It's not the same as having a 256-bit bus, but it's not nearly as bad as only having a single 128-bit one. With good planning and management of resources, more bandwidth can be had. It also means that the system has pretty good bandwidth for keeping the GPU fed by the SPEs, something that might not be possible if the system had stuck with a slower path between GPU and CPU just to get more GPU->vram bandwidth.

That's Blue. This is the most important information about the RSX and the PS3 that has been uncovered for the better part of this calender year.

Why haven't more people actually cared to talk about this? Gofreak what advantages does this mean for games like Warhawk that procedually creates the ocean on the CELL and passes that data to RSX for rendering?

And not to start a systems war but does the Xbox 360 have extra split memory like this? I'm only asking to gain truth. NO SYSTEMS WAR!!

 

[BlueTsunami]

Wow I admire you started, actually. On my side, I even couldn't do it...

:lol

I orginaly registered to see the nice scans they post but it became to much to try and filter out the BS and the legit stuff. Its more hassle than its worth and I figured it would be best if I lurk there (on rare occasions)

 

[Tat]

That's Blue. This is the most important information about the RSX and the PS3 that has been uncovered for the better part of this calender year.

Why haven't more people actually cared to talk about this? Gofreak what advantages does this mean for games like Warhawk that procedually creates the ocean on the CELL and passes that data to RSX for rendering?

And not to start a systems war but does the Xbox 360 have extra split memory like this? I'm only asking to gain truth. NO SYSTEMS WAR!!

Basically the 360 is like one single island. The PS3 on the other hand is like a number of islands connected by bridges. If put together the PS3 islands are bigger than the one 360 island though.

Nah'm saying?

 

[mckmas8808]

Basically the 360 is like one single island. The PS3 on the other hand is like a number of islands connected by bridges. If put together the PS3 islands are bigger than the one 360 island though.

Nah'm saying?

Yeah I feel you son, but I just can't figure out what this means for actual games. It would seem to me that people all over the internets would point this out as an advantage. Yet nobody really cares.

Is it because it's not really a big advantage after all? :/

 

[gofreak]

If you're asking what the advantage of two pools of memory over one are, I guess two busses = less contention, and a local pool for each chip means most of their accesses should be lower latency (versus one having to always have higher latency accesses).

360, you could consider as having two pools of memory, but one is used only for certain accesses (framebuffer).

The inter-chip bandwidth is a slightly different thing. It's really so large, IMO, to facilitate RSX and Cell accessing each others memories liberally. In terms of Cell procedurally generating stuff or whatever, I'm slightly more interested by what the extra computational capacity brings versus the larger CPU<->GPU bandwidth, but having both is great, of course.

 

[aaaaa0]

And not to start a systems war but does the Xbox 360 have extra split memory like this?

PS3

XDR <--25.6GB/s--> CELL <--35GB/s--> RSX <--22.4GB/s--> VRAM

XBOX 360

                                    Xenon
                                      |
                                   22.4GB/s
                                      |
GDDR3 <--22.4GB/s--> Xenos <--32 GB/s--> Xenos (ROPS)<--256GB/s-->VRAM

 

[mckmas8808]

PS3

XDR <--25.6GB/s--> CELL <--35GB/s--> RSX <--22.4GB/s--> VRAM

XBOX 360

                                    Xenon
                                      |
                                   22.4GB/s
                                      |
GDDR3 <--22.4GB/s--> Xenos <--32 GB/s--> Xenos (ROPS)<--256GB/s-->VRAM

Thanks but I meant a split system bus with the GPU. It's now said that the PS3 has 2 128-bit busses. One to CELL and the other to VRAM. To me and that PS3 developer it seems to an advantage. I was asking if the Xbox 360 was setup in a similar way.

It doesn't seem like it is.

 

[aaaaa0]

Thanks but I meant a split system bus with the GPU. It's now said that the PS3 has 2 128-bit busses. One to CELL and the other to VRAM. To me and that PS3 developer it seems to an advantage. I was asking if the Xbox 360 was setup in a similar way.

It doesn't seem like it is.

Xenos is the memory controller for the system. It has 3 busses.

One to the CPU @ 22.4GB/s.
One to GDDR3 @ 22.4 GB/s.
One to the ROPs @ 32 GB/s (and depending on how you count, one more from the ROPs to the VRAM @ 256GB/s).

The GPU can use all 3 at the same time.

 

[gofreak]

Thanks but I meant a split system bus with the GPU. It's now said that the PS3 has 2 128-bit busses. One to CELL and the other to VRAM. To me and that PS3 developer it seems to an advantage. I was asking if the Xbox 360 was setup in a similar way.

It doesn't seem like it is.

360 has an advantage for framebuffer bandwidth, if solely looking at what RSX can access (versus, say, the bandwidth at Cell's disposal if it was also making a contribution). PS3 for anything else, really.

The split memory setup in PS3 isn't really news though

 

[mckmas8808]

Xenos is the memory controller for the system. It has 3 busses.

One to the CPU @ 22.4GB/s.
One to GDDR3 @ 22.4 GB/s.
One to the ROPs @ 32 GB/s (and depending on how you count, one more from the ROPs to the VRAM @ 256GB/s).

The GPU can use all 3 at the same time.

At what bit? Just curious.

 

[aaaaa0]

360 has an advantage for framebuffer bandwidth, if solely looking at what RSX can access (versus, say, the bandwidth at Cell's disposal if it was also making a contribution). PS3 for anything else, really.

The easiest way to describe it is that the 360 is a GPU-centric design. The GPU gets the lion's share of the bandwidth in the system, simply because graphics is (and is intended to be) the largest bandwidth consumer.

 

[aaaaa0]

At what bit? Just curious.

The GPU can be pulling geometry from the Xenon's L2 cache, at the same time it's pulling textures from GDDR3, at the same time it's reading Z and writing pixels to the framebuffer.

3 busses.

 

[Tat]

XBOX 360

                                    Xenon
                                      |
                                   22.4GB/s
                                      |
GDDR3 <--22.4GB/s--> Xenos <--32 GB/s--> Xenos (ROPS)<--256GB/s-->VRAM

Isn't that 256GB/s only to the eDRAM?

 

[gofreak]

The easiest way to describe it is that the 360 is a GPU-centric design. The GPU gets the lion's share of the bandwidth in the system, simply because graphics is (and is intended to be) the largest bandwidth consumer.

Framebuffer-orientated, I'd say, which is one bw consumer of a couple for the GPU (if an intensive one, of course).

 

[aaaaa0]

Isn't that 256GB/s only to the eDRAM?

The eDRAM is the primary framebuffer VRAM for the GPU. 10MB is sufficient enough to tile such that the bandwidth drain on main memory is minimized.

 

[Tat]

The eDRAM is the primary framebuffer VRAM for the GPU. 10MB is sufficient enough to tile such that the bandwidth drain on main memory is minimized.

Yeah I forgot 360 has 512 GDDR3 shared.

 

[aaaaa0]

Framebuffer-orientated, I'd say, which is one bw consumer of a couple for the GPU (if an intensive one, of course).

No, I'd definitely call it GPU-centric (not just framebuffer-oriented), because all system busses go through the GPU, and the GPU gets to use bandwidth from anywhere in the system (GDDR3, VRAM, L2) without going through a middleman.

 

[mckmas8808]

The GPU can be pulling geometry from the Xenon's L2 cache, at the same time it's pulling textures from GDDR3, at the same time it's reading Z and writing pixels to the framebuffer.

3 busses.

I'm not hardcore so educate me a little. This might sound stupid but do you split the data that goes over that 128 bit bus by 3? Would it be 42.6 bits per? Or will it be 128MB per?

 

[gofreak]

No, I'd definitely call it GPU centric (not just framebuffer-oriented), because all system busses go through the GPU, and the GPU gets to use bandwidth from anywhere in the system (GDDR3, VRAM, L2) without going through a middleman.

From that point of view, yes - all roads lead to Rome, so to speak. But if comparing the single bus to a two-bus system with twice the bandwidth, perhaps, the latter would have been less tight for everything but the framebuffer.

 

[aaaaa0]

I'm not hardcore so educate me a little. This might sound stupid but do you split the data that goes over that 128 bit bus by 3? Would it be 42.6 bits per? Or will it be 128MB per?

No there are three seperate full bandwidth busses. You don't split anything. The bit-ness of a bus doesn't really matter, it's the bandwidth you care about.

 

[aaaaa0]

From that point of view, yes - all roads lead to Rome, so to speak. But if comparing the single bus to a two-bus system with twice the bandwidth, perhaps, the latter would have been less tight for everything but the framebuffer.

Why are you still going on about a single-bus?

The GPU has 3 busses, one to GDDR3, one to the CPU, and one to the ROPS.

 

[gofreak]

Why are you still going on about a single-bus?

The GPU has 3 busses, one to GDDR3, one to the CPU, and one to the ROPS.

I'm talking about memory bandwidth. The CPU bus is mostly going to be talking to and from memory. The ROPs bandwidth I've always acknowledged.

 

[mckmas8808]

Why are you still going on about a single-bus?

The GPU has 3 busses, one to GDDR3, one to the CPU, and one to the ROPS.

See that's what's confusing me. Gofreak (a person that I've trusted in information for a long time now) is stating the X360 is a single bus system, while the PS3 is not. Damn we need some Xbox 360 devs like element to confirm so information.

 

[aaaaa0]

I'm talking about memory bandwidth. The CPU bus is mostly going to be talking to and from memory. The ROPs bandwidth I've always acknowledged.

If you're talking about main memory bandwidth, talk about main memory bandwidth.

Don't say single bus system, because it's not a single bus system -- it's a 3 bus GPU-centric design.

 

[aaaaa0]

See that's what's confusing me. Gofreak (a person that I've trusted in information for a long time now) is stating the X360 is a single bus system, while the PS3 is not. Damn we need some Xbox 360 devs like element to confirm so information.

Just read the Beyond3D article that's been posted a bunch of times already.

http://www.beyond3d.com/articles/xenos/

 

[gofreak]

If you're talking about main memory bandwidth, talk about main memory bandwidth.

Don't say single bus system, because it's not a single bus system -- it's a 3 bus GPU-centric design.

Sorry for the confusion. In the context of mckmas's question, he's quite clearly asking about one bus to memory vs multiple busses to memory, so that's the angle I'm coming from, that's what I thought we were talking about.

To answer that question, the advantage would be in contention and latency for one of the chips, if comparing to a single bus to memory even with the same bandwidth as the two memory busses in PS3 combined. Disadvantage is that you will want to be more strategic about where different resources go.

And as I said earlier, if comparing to 360, it's not fair to ignore the eDram which is another bus to another pool of memory, which will take framebuffer-related contention/activity virtually wholly away from the main memory bus.

 

[aaaaa0]

The CPU bus is mostly going to be talking to and from memory.

That completely depends on what mode you're using the CPU in. If you lock L2, and just use the cores for procedural geometry generation, you can have very little main memory traffic, for at least some number of cycles.

In some future, I can envision phases of rendering in which the CPU does nothing but geometry generation directly out of L2, which would allow the GPU to suck up practically the full bandwidth of the system for at lest a little while. (Similar techniques for CELL of course.)

Anyway, real world performance isn't as simple as adding up one bus, two busses, or three busses and getting a single number out of it.

 

[gofreak]

That completely depends on what mode you're using the CPU in. If you lock L2, and just use the cores for procedural geometry generation, you can have very little main memory traffic, for at least some number of cycles.

For some number of cycles would be the key bit I think typically the cpu->xenos bandwidth will be mostly occupied with memory access, and any bandwidth used in such a fashion is not additive with main memory bandwidth in terms of what's going in or out of xenos. Even with procedural geometry, if it is to be persistent beyond the current frame, it'll have to hit main memory at some stage (unless you want to recompute it every frame!).

 

[Shompola]

I'm not hardcore so educate me a little. This might sound stupid but do you split the data that goes over that 128 bit bus by 3? Would it be 42.6 bits per? Or will it be 128MB per?

To that GDDR3 thing it should be 224 00000000 /1400 000000 * 8 = 128 bit
To the edram thing it should be 32 000000000 / 500 000000 * 8 = 512 bit
internal edram I dont know.

 

[aaaaa0]

(unless you want to recompute it every frame!).

I think that's the idea. You trade memory and bandwidth for CPU cycles, since (at least on paper), CPUs like CELL and Xenon are so rich for FLOPS and poor on main memory bandwidth. (IE you take some sort of higher-order representation, tesselate it on the fly and render that.)

Of course it hasn't worked out that way too well in the past, but who knows, maybe this time.

 

[ionicbluebird]

Doesn't the RSX basically have 47 GB's of bandwidth available... 22.4 to DDR, and 25.6 to XDR?

Xenos, basically has 54 GB's... 32 GB's to the ROP's, and 22.4 to the DDR.

From where I'm sitting, it would seem that the EDRAM's main benefit is in sustained fillrate of 4 gigapixels.

The biggest difference in IQ between the systems could be 2xaa vs. 4xaa.

 

[mckmas8808]

Doesn't the RSX basically have 47 GB's of bandwidth available... 22.4 to DDR, and 25.6 to XDR?

Xenos, basically has 54 GB's... 32 GB's to the ROP's, and 22.4 to the DDR.

From where I'm sitting, it would seem that the EDRAM's main benefit is in sustained fillrate of 4 gigapixels.

The biggest difference in IQ between the systems could be 2xaa vs. 4xaa.

So in your opinion what is Warhawk on the PS3 giving up for it's 4xaa?

 

[jax (old)]

It's funny how in 5 minutes people can debunk what the Xbox crew at OA has been preaching for the better part of a year. :lol

Thanks!

they post the same thing over and over daily and go stir crazy regarding PS3vsxbox360. Deadmeat + redcloak + reanimated + some of the other members needs to be taken out to the paddock and shot.



this is good news too. Can't wait for EEE to see what SONY brings to the table

 

[Raven.]

So in your opinion what is Warhawk on the PS3 giving up for it's 4xaa?

2xaa

 

[Leatherface]

 

[SolidSnakex]

Yes SoE sucks when it comes to graphics.

 

[ionicbluebird]

So in your opinion what is Warhawk on the PS3 giving up for it's 4xaa?

Well, in my opinion, RSX will have enough bandwidth for 4xaa at 720p, but as far as I know it will decrease the available fillrate to about roughly half of what no aa would give you.... based on the 7900 of course.

Of course that's still a lot of fillrate.

Anyone else with more knowledge wanna chime in?

 

[jax (old)]

i could swear that posting this screen for the reasons you are = trolling. this is still the physics thread



seriously.

 

[Tat]

Good thing every PS3 game will look like that just like every 360 gamed looked like the first wallguy screens.

Oh wait...

 

[Ryudo]

Good thing every PS3 game will look like that just like every 360 gamed looked like the first wallguy screens.

Oh wait...

If only the 360 games looked as good as the wallguy screens you would have a point. Slim pickings for the 360 thats for sure, it will make it all the more interesting to see the initial PS3 stuff.