Rumble Roses developer on PS3 & X360

[Nite_Hawk]

We also have no idea what other changes they have made. Will it be a straightup G70, or something more like the xbox gpu when compared to the geforce2?

Btw Shifty, thank you for pointing out the context of Deano's quote. I never knew that, and actually was a bit suprised he had said it. That makes a lot more sense.

Nite_Hawk

 

[Lysander]

And?

Jawed

1) if cell can tap rsx-ram and rsx cell-ram, then there will be a lot of data traffic in ps3 and possible bottlenecks
2) if not, then cell will be limited in graphic creation as the whole geometry-texture data should be stored on two separate memory hubs

 

[Jawed]

Cell has 300GB/s of internal bandwidth to shuffle stuff around. Bandwidth from one side of Cell to another is most definitely not a problem

Jawed

 

[Shifty Geezer]

Plus geometry can be fed straight to RSX without touching either RAM pool/bandwidth (an architectural feature the XB360 shares, I add with my amazing knowledge on the subject 'coz I'm just sooo clever and always know what I'm talking about, no really, no honest guv )

 

[Lysander]

I am not talkin about internal BW; I am talking about cell-rsxram, cell-cellram, rsx-cellram, rsx-rsxram, cell-rsx data stream.

 

[Titanio]

I am not talkin about internal BW; I am talking about cell-rsxram, cell-cellram, rsx-cellram, rsx-rsxram, cell-rsx data stream.

There is a lot of bandwidth between all of these components. Both RSX and Cell have more incoming bandwidth from Flexio than could be filled by the memory at the "other end". That won't be a problem, I don't think.

 

[Jawed]

Plus geometry can be fed straight to RSX without touching either RAM pool/bandwidth (an architectural feature the XB360 shares, I add with my amazing knowledge on the subject 'coz I'm just sooo clever and always know what I'm talking about, no really, no honest guv )

Fingers-crossed that it actually turns out that way...

Actually I can't imagine it not - that would just be rediculous.

Jawed

 

[Lysander]

That won't be a problem, I don't think.

really, because identical issue at X2 (which demands less inter connectivity than ps3 because unified memory) creates a lot of critical talk here

 

[Titanio]

really, because identical issue at X2 (which demands less inter connectivity than ps3 because unified memory) creates a lot of critical talk here

Sorry, I'm a little confused. Can you be more specific?

 

[Lysander]

Well, my understanding is that both memory chips are separate entity so if cell will need data for texture and the data will be stored in rsx memory hub, cell will need to tap that memory thru rsx data stream path therefore interfering with rsx demand; or perhaps cell will make a picture and save the picture in its own memory chip, creating the need for rsx to "go" for that picture data in "foreign" memory
This is the same problem as with X2 "bottleneck" issue here: connections X2cpu-X2gpu, X2cpu-(thru X2gpu)-memory, X2gpu-memory. Because X2cpu is getting info thru front bus from X2gpu AND unified memory (which is plugged on X2gpu), that create a bottleneck issue which is identical to cell-rsxram issue.

 

[darkblu]

Lysander, from all we know both chips within ps3 are not more separate than they'd be in a typical NUMA design. and if we step on that we can further assume that:
1) each chips has a reasonably direct access to non-local (i.e. peer's) memory*, and
2) data that absolutely needs random access from both chips should get mirrored into the respective local memory pools.

practice has shown the above two to actually work not bad under the conditions of comparably-demanding processing elements, as in the PS3 case with cell and rsx.

* meaning "not-that-much-penalized bandwidth" and "reasonably-predictably-penalized access latency", hence favorable conditions for streamed data access.

 

[Lysander]

data that absolutely needs random access from both chips should get mirrored into the respective local memory pools

.
Can I understand this as a duplication of same data? Won`t that take away to much memory space? (with notion that cell should do demading graphic performance acts)

 

[darkblu]

yes, but then again data that need random, concurrent access from both peers is rare to occur, due to the facts that A) under NUMA you don't scatter arbitrarily-used data randomly all over the mem pools, and B) that the peers here are not two cpu's but a cpu and a gpu. so do not expect too much data duplication due to localization, although yes, it may occur to a limited degree.

 

[Titanio]

Well, my understanding is that both memory chips are separate entity so if cell will need data for texture and the data will be stored in rsx memory hub, cell will need to tap that memory thru rsx data stream path therefore interfering with rsx demand; or perhaps cell will make a picture and save the picture in its own memory chip, creating the need for rsx to "go" for that picture data in "foreign" memory
This is the same problem as with X2 "bottleneck" issue here: connections X2cpu-X2gpu, X2cpu-(thru X2gpu)-memory, X2gpu-memory. Because X2cpu is getting info thru front bus from X2gpu AND unified memory (which is plugged on X2gpu), that create a bottleneck issue which is identical to cell-rsxram issue.

The difference with X2 is that the CPU will always "interfere" with the GPU's access, and vice versa. They're both always sharing the same pipe.

 

[Shifty Geezer]

.
Can I understand this as a duplication of same data? Won`t that take away to much memory space? (with notion that cell should do demading graphic performance acts)

There was a patent from Sony to avoid data duplication. Something about caching and tiling and magical stuff of wierd graphics super-capabilitiness.

I'll repeat (something you seem to be ignoring) that Cell is capable of passing data DIRECTLY to RSX. It can work in XDR, send data to RSX, which processes it and stores whatever it needs. Both chips can also access each other's memory pool, though not as quickly as they can access their own pool. PC's have been getting by with seperate RAM and VRAM for years so I don't see why you have a problem with this arrangement considering the other optimizations added to PS3's data pathways. Granted UMA of 512 mb XDR at 50 GB/s would have been nice, but there's always improvements that can be made if you don't care about cost.

 

[Titanio]

Granted UMA of 512 mb XDR at 50 GB/s would have been nice, but there's always improvements that can be made if you don't care about cost.

Nicer than what we have though? From a dev point of view, perhaps, that would indeed be easier. But following a X360 model with that, then all of Cell's access would be like it is now when it accesses RSX's memory. And given how latency-sensitive CPUs are, I'm not sure if that would be desireable. The plus point would be that all the GPU's memory would be local to it, but since GPUs (apparently) are not as latency sensitive, I'm not sure how much of a win that would be. I guess you could also cut down on your bw requirements between the two chips, but since we have so much with this design already..

 

[darkblu]

Granted UMA of 512 mb XDR at 50 GB/s would have been nice, but there's always improvements that can be made if you don't care about cost.

actually no, as the nondeterministic latencies would hurt your performance seriously (see xbox for a glaring example). i suggest you also google and read on NUMA vs UMA topologies.

 

[Shifty Geezer]

Ack! I'm not even going there! Yes I guess latencies would go up. Both NUMA and UMA have pro's and cons (shock, horror ) and neither is a perfect solution (gadzooks! Who'd've thought it! ) and I really don't appreciate the pros and cons either as I'm not writing for either console!

 

[darkblu]

shifty, not going where? my post had nothing to do with your appreciation or lack thereof of uma or numa. it was a direct comment to one statement you made (which i quoted). what you read into that so it provoked that reaction in you is beyond my undertanding :shrug: if that was the suggestion to google for information - my apologies, that was sincere and did not mean to put you down in any way, it's just that the information _is_ readily available through google.

 

[Shifty Geezer]

I took no offense Just saying I'm not going to talk about whether PS3 would be better off with UMA or not. Like all these things there's pros and cons, and no definitive right or wrong. Seeing as all these points have been rasied already on this forum I can't see much reason to restart talking about whether 512 MBs XDR > 2x256 MBs XDR+GDDR.

But if others wanna talk about it that's fine. Just not a topic I personally want to go over again.