"PS3 Reference Tool" Ram Type Overlooked?

[ROG27]

Way back when at the PlayStation Meeting, Sony released slides of the dev kits which were to roll out in the coming months. I find it interesting that the final 2 iterations have 512 MB of XDR RAM, not 256 XDR and 256 GDDR3 like the intial specs stated. This upgrade would really kick ass. Could this be something most of us simply overlooked?

 

[Carl B]

It wasn't overlooked, I think it was just assumed 512MB XDR for 'main' memory, 256MB GDDR3 still for the RSX. For example Spring 2005 512 XDR doesn't make any sense in the context of RSX, since RSX isn't ready(?) yet. And you know the GTX or Ultra or whatever isn't using XDR.

 

[nAo]

Historically dev kits have more memory than retail units (PS2 devkit has 4x times the memory of the 'real' PS2), so there's anything new here..

 

[Mefisutoferesu]

How does CELL address the extra RAM? I thought the upper limit was 256MB, or is my memory failing me and that was debunked?

 

[pakpassion]

and boom! there goes another hope(fullness)

 

[Carl B]

How does CELL address the extra RAM? I thought the upper limit was 256MB, or is my memory failing me and that was debunked?

Yeah that was debunked. Second half of the page I'm linking to for the explanation:

http://realworldtech.com/page.cfm?ArticleID=RWT022805234129

 

[Thowllly]

How does CELL address the extra RAM? I thought the upper limit was 256MB, or is my memory failing me and that was debunked?

There is no such upper limit. Some have incorrectly assumed that you can at most use 4 memory chips with the cell (and with 64MB chips that's 256MB), but you can in fact hook up at least 32 memory chips to the cell (or up to 36 with ECC). And I don't think there's any reason to believe that the largest chip size you can use is 64MB either.

Edit: 36 chips is for one memory channel. The cell has two, so I should have said 72... not the first time I make that mistake...

 

[Mefisutoferesu]

Yeah that was debunked. Second half of the page I'm linking to for the explanation:

http://realworldtech.com/page.cfm?ArticleID=RWT022805234129

Ahem...

----------__________----------


Seriously, I need memory supplements or something... Sorry, for the annoyance everyone.

 

[Laa-Yosh]

Still, the Cell servers that have been announced recently only have 512MB of RAM per CPU...

 

[MBDF]

256 for cell, 256 for RSX?

 

[phat]

Still, the Cell servers that have been announced recently only have 512MB of RAM per CPU...

That's because while >4 chip addressing is certainly theoretically possible, the current incarnation of the Cell probably has an XDR interface that is hardwired for 4. (This would greatly simplify the design and validation of a high-speed interconnect like XDR.) Otherwise, it makes no sense that these Cell blades and clusters are limited to 512MB/Cell. That is a pitiful amount of memory for their intended applications.

 

[V3]

Still, the Cell servers that have been announced recently only have 512MB of RAM per CPU...

It a 64 bit CPU so it can address more than that. But the whole Cell architecture is like that, instead of one large pool of memory with lots of cores, its a multiple segmented memory, each segment shares by smaller number of cores.

 

[hugo]

With the introduction of GDDR4 and the reduction of the XDR cost you can't rule out the possibility of it.Sony may up their XDR size in the PS3 at the last minute if they wanted to.

 

[Carl B]

Since this thread has kind of by default become a 'Cell in the server/workstation space' thread, I remembered this tidbit from a while ago:

There are already some modification going on within CELL. There is some progress in replacing the entire FPU inside the SPU with a full-blown DP (double precision) unit. The estimated performance should be about 1:2 against the current SP unit, which would be a major improvement compared to the current situation of about 1:10 - 1:14.

Link

Not sure if it was ever posted around here or not, but no way I'm going to run a search for 'Cell'...

 

[PC-Engine]

With the introduction of GDDR4 and the reduction of the XDR cost you can't rule out the possibility of it.Sony may up their XDR size in the PS3 at the last minute if they wanted to.

I'm sure they want to bleed more money than they're currently.

 

[Gholbine]

I'm sure they want to bleed more money than they're currently.

It would help push them even further ahead of the competition graphically, and an extra 256mb of RAM is something no developer could refuse. It would help a lot, but it would likely cost a lot. I don't see it happening.

 

[Thowllly]

That's because while >4 chip addressing is certainly theoretically possible, the current incarnation of the Cell probably has an XDR interface that is hardwired for 4. (This would greatly simplify the design and validation of a high-speed interconnect like XDR.) Otherwise, it makes no sense that these Cell blades and clusters are limited to 512MB/Cell. That is a pitiful amount of memory for their intended applications.

The way XDR works, I don't see how it would be possible for it to be hardwired for any number of chips. XDR has one address bus that is output only, and all addresses goes to all the chips. The address bus works at 1/4 the speed (in transfers/s) of the data bus and is basically the same as the bus that's used RDram (for both data and address) except that it's only used as an output bus for addresses. The data bus is 36 bits wide (32 data + 4 ECC) and used for both input and output. Each of these data lines is connected to one chip and one chip only. There can be any number of chips as long as all 36 data lines (or 32 without ECC) are connected to one chip and one chip only. So you could use 7 chips that are 3 bits wide, plus one chip that's 11 bits wide for a total of 32. The only reason I can think of for the low amount of memory is because for now there only exists XDR chips that are 16 bits wide with 32M words, or something like that.

 

[Thowllly]

Since this thread has kind of by default become a 'Cell in the server/workstation space' thread, I remembered this tidbit from a while ago:



Link

Not sure if it was ever posted around here or not, but no way I'm going to run a search for 'Cell'...

Nice, I hadn't seen that before. I doubt that will make it's way into the PS3, since DP is not very important for games. On the other hand, if they want to have just one type of Cell for everything then maybe it will end up in the PS3...

 

[Guden Oden]

That's because while >4 chip addressing is certainly theoretically possible, the current incarnation of the Cell probably has an XDR interface that is hardwired for 4.]

That's not true. There are pics out there with at least 8 memory chips hooked up to the Cell.

(This would greatly simplify the design and validation of a high-speed interconnect like XDR.)

No, it would not. To get the particular memory size and bandwidth combination used in PS3 you need all the XDR bitlines of the Cell hooked up to memory devices. So the entire memory interface needs to work as intended. If you hook the bitlines up to four RAM chips total or 32 makes no technical difference due to the way XDR works.

 

[Carl B]

Nice, I hadn't seen that before. I doubt that will make it's way into the PS3, since DP is not very important for games. On the other hand, if they want to have just one type of Cell for everything then maybe it will end up in the PS3...

Yeah, I'm wondering about the same thing, but I guess the safe bet right now would be a DD3 revision for IBM's third-party customers. The die size would increase with the DP unit and the time to implement might be another strike against including such in the PS3. Still, I could understand them perhaps not wanting to have different Cell-lines for production this early in the game.