Are PS3 devs using the two mem pools for textures?

And if so, what is the way it´s managed the division ? For example, to postprocessing can it be send one frame to the GDDR3 pool, the next one to XDR and son on ? Help me to clear this subject.
 
And if so, what is the way it´s managed the division ? For example, to postprocessing can it be send one frame to the GDDR3 pool, the next one to XDR and son on ? Help me to clear this subject.
I doubt you'll find any dev that desperately wants to lose his/her job.
 
If you go back to sony's E3 press conference, Jen Huang made a point of stressing that all the memory could be utilized by the RSX. One of their slides states explicitly that the ps3 has 512mb of graphics render memory. According to Jen.

The RSX can render pixels to anywhere in system memory.
 
The RSX can render pixels to anywhere in system memory.

Can and can efficiently are 2 different things. Although I'm sure there are some restrictions, because, after all, NUMA and UMA are not the same, the RSX can clearly access and write to XDR fast enough to make it worth while. The question is, what are the penalties associated with it. I'm too lazy to google it, but I know the research has been done and this info is public knowledge. B3D I believe has done some coverage on this subject.
 
i think the subject is more if it is an easy but penalty thing or if it is difficult and requires an exhotic way of making it. If it is the last, couldn´t be added some kind of texture mem manager in a devs tool kit as edge to be used by most of the developers or not ?.
 
What was the benefit of going with a split pool, again? Or, more specifically, XDR?
It seems like everything related to the memory system is a way of getting around it, to the point where I wonder why they didn't go with a UMA or, if they wanted to keep two buses, two pools of GDDR3.
 
What was the benefit of going with a split pool, again? Or, more specifically, XDR?
Twice the bandwidth at a lower cost, and lower latency from the XDR. Two pools of GDDR would net you the same problems as you'd still have to go through another chips memory interface to access its pool of RAM, and the Cell would be paired with slower memory.
 
What was the benefit of going with a split pool, again? Or, more specifically, XDR?
If for some reason you're mem bandwidth limited while using only one bus you try to remove your bottleneck reading and/or writing data using multiple buses, it's just as simple as that.
 
What was the benefit of going with a split pool, again? Or, more specifically, XDR?
It seems like everything related to the memory system is a way of getting around it, to the point where I wonder why they didn't go with a UMA or, if they wanted to keep two buses, two pools of GDDR3.

Because in the beginning it was thought to be used 256 Mb of RAM. Rambus memory was signed many years before finishing the PS3 desing, when the A plan was to use two cells and no PC style GPU. After that in the last moments it was decided to use RSX and add more memory ( only because of Microsoft decission in increasing mem ? ). XDR is very expensive, so adding other 256 mb of this kind of memory was prohibitive, as well as RSX being better suited to conventional PC-GPU memory types ( the bridge between RSX and XDR was one of the last things to be designed in PS3 ).
 
If for some reason you're mem bandwidth limited while using only one bus you try to remove your bottleneck reading and/or writing data using multiple buses, it's just as simple as that.

That is the good thing now, once things are made, but do you think that if sony was now in the design table five years ago it would be like that again ?.
 
Last edited by a moderator:
The other choice was a 512 MB pool of XDR, which would need a GPU able to access it and a far higher price for the memory needing a larger bus.
 
Because in the beginning it was thought to be used 256 Mb of RAM. Rambus memory was signed many years before finishing the PS3 desing, when the A plan was to use two cells and no PC style GPU. After that in the last moments it was decided to use RSX and add more memory ( only because of Microsoft decission in increasing mem ? ). XDR is very expensive, so adding other 256 mb of this kind of memory was prohibitive, as well as RSX being better suited to conventional PC-GPU memory types ( the bridge between RSX and XDR was one of the last things to be designed in PS3 ).

Nope, there has to my knowledge NEVER been a 2 cell design PS3...
 
Nope, there has to my knowledge NEVER been a 2 cell design PS3...

ups, maybe toshiba´s gpu instead. The truth is out there, in you, sony´s related people.

By the way, i´ve just seen the last video in jeux videos from your game, and i must give you folks my congrats, whatever you´re doing with the PS3 bits and its mem pools is enormous!
 
To my understanding, the RSX can access both, and for a lot of games, does access both.

The bottleneck simply isn't there as some believe it to be, and if it were, it would be a big issue, which thus far, many PS3 games have shown us that it clearly is not.

However, it is also in my understanding that the CPU cannot read well, at all, from the PS3's GPU memory (GDDR3) and that it's transfer rate is something like 5 megabytes a second? It's pretty slow, at least. But then again, I can't remember the last time I'd seen the CPU want to *read* the GPU's memory, when the GPU can simply write the information the CPU needs, and then the CPU can simply fetch it there.
 
I think it was 6 mb/s writing, and about 60 mb/s (Still slow) reading directly between CELL and the GDDR3 ram.

As for the 2 CELL design, I believe it only existed as an idea in the first few moments of the PS3s design life and was quickly discounted.

It would've been unwise of Sony not to investigate the benefit/drawback of using one for graphics originally.
 
I think it was 6 mb/s writing, and about 60 mb/s (Still slow) reading directly between CELL and the GDDR3 ram.


PS3_memory_bandwidths.jpg
 
Just a little reminder: The figures need a little adjustment for the lowered clock (not that's it's that big of deal, but I am anal :D )
 
Because in the beginning it was thought to be used 256 Mb of RAM. Rambus memory was signed many years before finishing the PS3 desing, when the A plan was to use two cells and no PC style GPU. After that in the last moments it was decided to use RSX and add more memory ( only because of Microsoft decission in increasing mem ? ). XDR is very expensive, so adding other 256 mb of this kind of memory was prohibitive, as well as RSX being better suited to conventional PC-GPU memory types ( the bridge between RSX and XDR was one of the last things to be designed in PS3 ).

That's pure supposition on your part. For one thing, sony uses rambus memory in the ps2, so sony has had a relationship with Rambus for many years.

I remember reading a business tech article as early as summer 2002, shortly after MS's and nvidia's well publicized falling out over xbox chip pricing, where it was reported that Sony and Nvidia were to colloborate on the gpu for the ps3. What was interesting was sony's reaction to said article. Instead of issuing the customary "we don't comment on rumors or speculation", sony derided the report, labelling the concept of sony requiring any external party to co-develop a gpu as ludicrous, since sony had designed the gpus for both the ps and ps2. Fastforward two years later and we get the official confirmation that nividia is co-designing the gpu along with sony and the revelation that they'd been at it for two years. So, the notion that the RSX is some type of last minute bolt on addition is not supported by facts or for that matter, by common sense.

You can check out the following link from 2004.

http://www.xbitlabs.com/news/multimedia/display/20041207105556.html

Jen-Hsun Huang, NVIDIA’s CEO and President said the two companies has worked closely “over the past two yearsâ€￾ on the “next-generation computer entertainment systemâ€￾.
 
Back
Top