so PS3's RSX ~ Reality Synthesizer GPU has *no* eDRAM ????

[DemoCoder]

X360GPU doesn't have eDRAM on core (its on a separate chip). It has 48gb/s bandwidth, which is not really mind bogglingly high. Why? Because there's no way they were gonna fit 10mb of eDRAM on the core and have good yields.

RSX has no eDRAM. Inquirer is full of it. And it wasn't for lack of time. an RSX with any significant eDRAM would be unfabb-able at 90nm. Maybe one day we will find out the full history of the ill-rated Sony designed GPU for PS3.

 

[Sonic]

Do we even know the cache structure of the GPU?

It's possible Sony has found a way to counteract needing all that bandwidth for the GS.

 

[Acert93]

I would not worry about the eDRAM much... the RSX looks to be really powerful and a lot of bandwidth--plus a lot of memory. It is about tradeoffs, and it seems the RSX went with more logic than eDRAM. There could be pros and cons, but overall I would not be dissapointed. It is, after all, about total system balance. And the PS3 has power and bandwidth oozing from it.

 

[loekf2]

According to the inquirer RSX is G70 plus 10MB EDRAM: http://www.theinquirer.net/?article=23325

BIG SIGH....

Except that our friend Fudo still doesn't get it.. Xbox 360's R500 is a R600 derivative and will have an unified shader architecture. R520, the new desktop part, is "just" a souped up R4xx GPU with more pipes and shader 3.0 support.... so still seperate vertex and pixel shaders.

 

[AlNom]

R500 is a R600 derivative

integrative....is that a word?

Anyways, it's the otherway around.

 

[gamepower]

X360GPU doesn't have eDRAM on core (its on a separate chip). It has 48gb/s bandwidth, which is not really mind bogglingly high. Why? Because there's no way they were gonna fit 10mb of eDRAM on the core and have good yields.

RSX has no eDRAM. Inquirer is full of it. And it wasn't for lack of time. an RSX with any significant eDRAM would be unfabb-able at 90nm. Maybe one day we will find out the full history of the ill-rated Sony designed GPU for PS3.

Why seperate?? I think it has 256Gbyte/s speed not 48Gbyte/s

 

[hugo]

According to the inquirer RSX is G70 plus 10MB EDRAM: http://www.theinquirer.net/?article=23325

Adding an EDRAM cache to the RSX would make the PS3 near flawless in design.

 

[jvd]

According to the inquirer RSX is G70 plus 10MB EDRAM: http://www.theinquirer.net/?article=23325

Adding an EDRAM cache to the RSX would make the PS3 near flawless in design.

it would . Except i'm sure it will be close to 400mt which i think would make yields so bad that two people will own a ps3 next year

 

[cthellis42]

Why seperate?? I think it has 256Gbyte/s speed not 48Gbyte/s

Microsoft indeed says 256GB, so even if separate it will still be fuggin' fast.

 

[DemoCoder]

If you use Microsoft's bandwidth counting method, than the RSX has 160gb/s, or more. And the PowerVR chip blows everyone away. 256gb is "effective" bandwidth, it is not real, and depends on compression and framebuffer mode. (e.g. no color compression with HDR rendertargets)

 

[ERP]

If you use Microsoft's bandwidth counting method, than the RSX has 160gb/s, or more. And the PowerVR chip blows everyone away. 256gb is "effective" bandwidth, it is not real, and depends on compression and framebuffer mode. (e.g. no color compression with HDR rendertargets)

In MS's case it isn't like traditional framebuffer compresson.
It has no worst case scenario, even if the traditional FB compression would provide no benefit, the EDRAM solution provides the same bandwidth saving as it does in it's best case.

 

[cthellis42]

If you use Microsoft's bandwidth counting method, than the RSX has 160gb/s, or more. And the PowerVR chip blows everyone away. 256gb is "effective" bandwidth, it is not real, and depends on compression and framebuffer mode. (e.g. no color compression with HDR rendertargets)

...in which case we really don't know what is "real" right now anyway, since not enough is known about their GPU yet. We're currently just comparing paper numbers, but how does identifying it as 48GB/s now (while all the other numbers we're dealing with are still "paper numbers") doing it a service? How did you reach that result, and how confident are you that the 360's GPU will be using the process you expect?

 

[Riddlewire]

Does Microsoft intend to reveal any further information about the X360 GPU during E3? I saw this quote from them in an interview today:

You take a look at the embedded DRAM, for example. We put over a million gates in the embedded DRAM that do processing internal to the embedded DRAM. To do things that would normally be done on the GPU side. It helps with some of the stuff that happens during the rendering process, which means you don't have to shift that stuff back and forth across the memory bus.

That would seem to at least confirm that it's a separate chip (which I think everybody realized anyway. But nobody seems to be able to figure out a good way to estimate the thing's performance.

 

[DemoCoder]

If you use Microsoft's bandwidth counting method, than the RSX has 160gb/s, or more. And the PowerVR chip blows everyone away. 256gb is "effective" bandwidth, it is not real, and depends on compression and framebuffer mode. (e.g. no color compression with HDR rendertargets)

In MS's case it isn't like traditional framebuffer compresson.
It has no worst case scenario, even if the traditional FB compression would provide no benefit, the EDRAM solution provides the same bandwidth saving as it does in it's best case.

But the benefit only comes with AA, and potentially no-HDR, so the worst case is lots of render-to-texture operations which don't use MSAA or backbuffers which use HDR. In practice, one has to ask is, for RSX's compression, what's the average case effectiveness, vs pathological case. It doesn't make sense to compare pathological cases, its more sense to look at what the expected case will be, and the expected workload.

 

[ERP]

Oh I'll agree with that.

All I'm saying is that MS's 256GB/s isn't some made up best case number based on variable compression, it's as real a number as any of them are. With 4xAA alpha blending and Z/Stencil compare/update that is the number you will get, which happens to match nicely the exact requirements of an 8 pipeline part with 32 bit color.

The bandwidth on the Sony/NV part concerns me (especially if they are serious about 1080), but I'll wait until I see it before I make judgements.

The ATI part has a few interesting new features, some targeted at HDR, some from WGF, and some to work around the restrictions of the EDRAM and I'm sure the NV part is similar in that respect.

It'll be interesting to see how both shape up in the end, I wouldn't want to declare a winner based on what's been disclosed thus far.

 

[Acert93]

The ATI part has a few interesting new features, some targeted at HDR, some from WGF, and some to work around the restrictions of the EDRAM and I'm sure the NV part is similar in that respect.

When you come out from NDA please share Specifically the above

 

[mech]

Doesn't the whole "embedded" name imply it's on-chip?

If it were a separate chip, that's effectively just more RAM - and I truly doubt they can get external ram running at that sort of speed. It's the short path lengths of embedded ram that lets it be run so fast.

Why are people saying it's an external chip? Got any sources to back you up there?

 

[Jawed]

Mech, it's an embedded DRAM chip because the chip has circuitry on it to do pixel blending, z-test and AA filtering.

This circuitry, in a normal PC GPU, consumes hideous amounts of bandwidth.

XB360's solution is to partition this functionality onto a separate chip. This means that not only can the GPU carry on doing other things, but the main memory being used by the GPU gets extra memory bandwidth for non "frame buffer" tasks.

Jawed

 

[Glonk]

The 360's AA is done via circuitry on the EDRAM chip...?

 

[Jawed]

Not entirely, but partially, yes.

Feel free to consult the patents:

http://patft.uspto.gov/netacgi/nph-Parser?patentnumber=6614449

http://patft.uspto.gov/netacgi/nph-Parser?patentnumber=6873323

The memory bandwidth issue can become especially troublesome when the bandwidth over a bus structure connecting a memory to a graphics processor is limited. In three-dimensional (3D) graphics processing, stored graphics data is often retrieved from memory and compared with newly generated pixel fragments to determine whether or not the fragments modify the state of the image as it is currently stored. If a determination is made that the fragment modifies the stored pixel information, the new value must be written back to the memory. As such, a large amount of bus bandwidth is required in order to perform both reading and writing operations associated with transferring video graphics data to and from the memory, which is often referred to as a frame buffer.

Therefore, a need exists for an anti-aliasing technique that makes efficient use of memory resources such that less memory space is required for storing the samples, and less bus bandwidth is required in order to perform the blending operations required in 3D graphics applications.

Jawed