GTX512/RSX analysis

[SynapticSignal]

I'm talking about..split internal bandwidth.

The bandwidth all to the GPU is like 22 GB or whatever.

Can it go into XDR for more? If so, dont you have to split the rendering engine or something? You have two physically seperate pools of 256 RAM. Will the latency difference be seamlessly overcomable? Is that even possible?

I'm not talking about CPU>GPU bandwidth. That's like what would go over PCI Express or AGP in a PC. It would be the 22 GB from Cell. But can the GPU use that like internal bandwidth?

I guess to my mind, PS3 is like a PC with 256 of system RAM, with a graphics card with 256 of video RAM.

Here is a crude diagram:

PC 7800 card>>38 GB

RSX>>22 GB>256GDDR
+>>22 GB >256 XDR

Total RSX has 44 GB. But into TWO smaller pools of RAM and introducing TWO differing latency levels, not one.

no
the XDR can't be addressed directly by rsx, there'are some big latences in this process, and others (big) with the XDR nature
it's like a kind of "turbo caches" of cheap nvidia cards, or (hyper memory) of ati

main local memory are 256 MB of 128 bit GDDR3

I think that this put in flames the whole gpu, correct me if I'm wrong, but aa and hdr eats a big amount of bandwidh working in the framebuffer, rsx don't have 256 bit memory or eDRAM, how can manage this effects?
my doubts came from the general low efficiency of g70 with aa and his severe performance hit, how can it be with a 128 bit bus?
(and no, you can't accumulate different buses when working on Framebuffer)

 

[nAo]

the XDR can't be addressed directly by rsx, there'are some big latences in this process, and others (big) with the XDR nature
it's like a kind of "turbo caches" of cheap nvidia cards, or (hyper memory) of ati

I don't understand: if XDR can't be directly addressed by RSX, then why are you saying it's like turbo cache where NVIDIA/ATI GPUs can address system memory...?

 

[SynapticSignal]

I don't understand: if XDR can't be directly addressed by RSX, then why are you saying it's like turbo cache where NVIDIA/ATI GPUs can address system memory...?

no, nvidia/ati can't do this directly
their gpu's data have to pass in the pcie-e slot interface (flexIO), north bridge in MB, then finally memory

I say "directly addressed", with the 256MB of local GDDR3 memory we have an high speed bus between gpu and memory, the difference in speed and latences is huge

 

[TurnDragoZeroV2G]

Then, realistically, what are they going to do with all that FlexI/O bandwidth? Isn't it supposed to be 35GB/s (for RSX, 20GB/s read and 15GB/s write), which, if 5GB/s were used strictly for CPU->GPU, would give a symmetrical 15GB/s in each direction. There should be higher latency going to XDR, but what exactly did Sony plan to use that BW for? Did they just get to the end and say, "Hey, you know, guys, we have some extra bandwidth through Cell's FlexI/O Let's use it for the GPU!" 15GB/s of write bandwidth... I don't see much point except using XDR as a backbuffer, making Cell do the backbuffer ops, or letting Cell do post-processing stuff. "Realtime Photoshop'd graphixxx!" (and actually, in that last case, wouldn't they be transmitting just the frontbuffer, which would be without Z and in 8 bits/channel?)

 

[SynapticSignal]

Then, realistically, what are they going to do with all that FlexI/O bandwidth? Isn't it supposed to be 35GB/s (for RSX, 20GB/s read and 15GB/s write), which, if 5GB/s were used strictly for CPU->GPU, would give a symmetrical 15GB/s in each direction. There should be higher latency going to XDR, but what exactly did Sony plan to use that BW for? Did they just get to the end and say, "Hey, you know, guys, we have some extra bandwidth through Cell's FlexI/O Let's use it for the GPU!" 15GB/s of write bandwidth... I don't see much point except using XDR as a backbuffer, making Cell do the backbuffer ops, or letting Cell do post-processing stuff. "Realtime Photoshop'd graphixxx!" (and actually, in that last case, wouldn't they be transmitting just the frontbuffer, which would be without Z and in 8 bits/channel?)

I think that flexIO was thinked to fetch some vertex data from cell'spu to rsx
it can't be used in heavy mission critical task as AA or HDR on the framebuffer, the devs will prefer of course the 256 MB local gddr3 to this tasks

My doubts are from the 128 bit bus between gpu and local memory, I'm pretty sure that this is not enough for the big performance hit of g70 (or rsx) structure when it comes to AA or HDR

I'll wait to see what devs will do..
I think that the launch titles for ps3 will lack hdr and aa, then the devs will found some workaround to this problem

 

[nAo]

I say "directly addressed", with the 256MB of local GDDR3 memory we have an high speed bus between gpu and memory, the difference in speed and latences is huge

Ok, now I see what you mean, but if the XDR mem protocol and FlexIO interface don't add a significantly data overhead I can't why overall bandwith should be affected if you design the GPU around the incresead latencies.
I think NVIDIA/ATI GPUs already try to address this problem (even if I know PCIE add a significantly protocol overhead..)

 

[Fafalada]

the difference in speed and latences is huge

Latency difference doesn't necesserily mean a speed difference. If the target hardware(or software) is equipped to hide asociated latencies, given the available bandwith of FlexIO, speed could be comparable to GDDR accesses.
It depends how hw was designed to handle this or not.

I'm pretty sure that this is not enough for the big performance hit of g70 (or rsx) structure when it comes to AA or HDR

I said this before - but usage patterns of recent PC games aren't exactly relevant benchmark to what targeted software will do on a closed platform.

 

[SynapticSignal]

Ok, now I see what you mean, but if the XDR mem protocol and FlexIO interface don't add a significantly data overhead I can't why overall bandwith should be affected if you design the GPU around the incresead latencies.
I think NVIDIA/ATI GPUs already try to address this problem (even if I know PCIE add a significantly protocol overhead..)

what I'm trying to say is that even the HM and TC cards have some little amount of local mem for the framebuffer (32/64 MB)
we can't think of using the flexIO per FB operations, if we don't want GF6200-like speed in AA, for what I say It can be done but putting huge latences in a speed mission critical tasks is a suicide for a dev
so we have a 128 local bus (suitable solution for FB tasks) AND flexIO (for vertex fetching or texture caching and so on)

that's why you CAN'T add 128 bit + flexIO when we talks of heavy bandwidh operations on the frame buffer
you can't work with two framebuffers, you have to choose if work in local or differed memory, and if you care a bit about performances, you have to forget the xdr, when coming to AA, HDR and blending
AFTER this effect are done, the FB can be moved in XDR in order to do some post processing via cell (I don't know the efficiency of moving the FB between the two banks of 256, we have to wait to understand if that is possible)

but stop the 22 + xyz GB/s when talking of aa, hdr, blending on frame buffer

because I don't believe that a 128 bit bus can be enought for this operations, I'll wait to see what devs will do to workaround the problem


[edited to correct some typos]

 

[nAo]

that's why you CAN'T add 128 bit + flexIO when we talks of heavy bandwidh operations on the frame buffer
you can't work with two framebuffer

Maybe you haven't noticed that I'm not adding XDR and VRAM bandwith regarding framebuffer operations. I've just told you that I can't see why there should be a significant impact on XDR available bandwith (from RSX) if this part is designed around the increased latencies. that's all!

but stop the 22 + xyz GB/s when talking of aa, hdr, blending on frame buffer

I'm sure you're talking about me here

 

[LunchBox]

Multiply those numbers by 100 million units and I beg to disagree...

not necessarily 100 million units...

first year would a smaller number maybe 10 Million?

the crazy boost on console sales occur after 2-3 years...

when it gets really affordable for the media...

around the 3rd or 4rth price drop...

and i agree to the rest of what you said

 

[Brimstone]

I would consider the AA and texture information new, but maybe that's just me. Anyway if you don't have anything to add Expletive, by all means don't.

I hope the AA and improved texturing information pans out.

As far as Faster GDDR3 goes, I'd rather see the 1Ghz clocked (8 Ghz effective octal signaling) XDR2 ram implemented on the RSX.

 

[nAo]

e poi c'era la marmotta che chiudeva la cioccolata..

 

[London Geezer]

e poi c'era la marmotta che chiudeva la cioccolata..

....







Cambridge is having a strange effect on this Italian...

 

[SynapticSignal]

e poi c'era la marmotta che chiudeva la cioccolata..

loooool
a marmotta with XDR at 40 GHz ( 800 GHz for each marmotta's finger)
OWNED

 

[Shifty Geezer]

Most of the things they've went with have been to stay competetive, they include 512MB of ram because to do otherwise would cost them market share, why not more? They went with the G70 derivitave because to adopt a more non-convential approach would've put them at a competetive disadvantage.
The only things Sony went "above and beyond" with also happen to be things that Sony has a vested interest in. CELL and BluRay are not in there to avoid the 'cheapest solution' they're in there because sony has a vested interest in these 2 technologies succeeding.

Dunno about that. We've got 6 USB ports, 3 different memcard slots, 3 Gigabit connections etc. Most of these aren't proprietary or does Sony have a vested interest in them, and most could probably be removed at a saving of several dollars, especially if they went with simpler alternatives. Obviously they're included as features to help promote the conosle. In the long run they'll cost Sony hundreds of millions of dollars over a cheapest possible minimum spec solution, but Sony have made the choice to include them as they feel the benefits outway the costs.

It seems if they can add a little extra functionality and performance at a measured cost, they will make a choice to or not. The argument 'Sony won't make a performance change that'll cost them more money' doesn't seem a correct evaluation. Cost isn't the only criteria. If it was half the connectors in PS3 wouldn't be there.

 

[Shifty Geezer]

Regarding what I felt the primary purpose of this thread, here's the excerpt from Josh's article with the rumours...

While nothing concrete is known about the 90 nm refresh of the G70, we can speculate that the product replacing the 7800 GTX will at least reach 550 MHz and that the high end could possibly hit 700 MHz at maximum. NVIDIA will not stop at just clock speed. We can expect that it will have a full 8 quads of pixel shaders, 10 vertex shaders, and 16 “Superâ€￾ ROPS that will be able to handle the output of those shaders. I have heard rumors that the AA unit will be getting a makeover and it will be able to handle HDR anti-aliasing. I have also heard rumors that texture filtering will also be getting a boost and we can expect texture quality to match that of the older FX series. This product could easily hit 380 million transistors, and with the addition of 90 nm Low-K (remember, the regular G70 is 110 nm FSG- it does not get a transistor performance increase by using Low-K) this product will hit some impressive clockspeeds. One thing that does not look to change will be the memory controller. NVIDIA does not feel the need for a programmable memory controller like ATI has as of yet, and will instead rely on faster GDDR-3 memory to make up the difference.

With nothing more concrete than 'a mate of mine knows a guy who was in the pub the other day where he overheard a girl saying her boyfrined was working on a 90nm refresh of G70 with improved AA' I can't see anything more can be added in our talk of RSX than usual. RSX might be getting HDR MSAA because it's rumoured. Then again it might not, because rumours often aren't true. C'est la vie

 

[nAo]

Then again it might not, because rumours often aren't true. C'est la vie

You're wise man

 

[SynapticSignal]

Dunno about that. We've got 6 USB ports, 3 different memcard slots, 3 Gigabit connections etc. .

I read somewhere that 3 usb and the gigabits are dropped of from the ps3 to contain the cost

 

[London Geezer]

I read somewhere that 3 usb and the gigabits are dropped of from the ps3 to contain the cost

Did you not get the memo?

Every rumour goes through ME!!!

Like this one...

or this one...

 

[SynapticSignal]

Did you not get the memo?

Every rumour goes through ME!!!

Like this one...

or this one...

hum yes
they have to do something to be competitive with 399 $ X360
if they sell ps3 for 500 $ or more, this will sell badly, very badly imo
boys in a range of 8-18 yo are not rich, usually and don't work so..
and a family father that want to buy a console with the family money, hardly will choose a 500+ $ console

so, sony will cut something, but...... what?
cell no, they can't
rsx no, they can't
memory no, they can't
3 usb? 'humm yes.. possible...'
br-dvd? 'humm, it's very expensive.. why don't... make 2 SKU...'
gigabit ports? 'oh yeah'