RSX Secrets

Status
Not open for further replies.
Cell:
65nm (212mm2 die size) technology

Are you sure about that? I thought it was 174.6mm2. 212mm2 might be the 65nm CELL with enhanced fp64 throughput, but I'm not sure about this...


RSX:
28 pixel shaders (4 redundant, 24 active)

Where does this come from? I've only seen specualtions about that, but nothing to suggest it's actually true. And if it's true there would be 28 texture units too.

Rambus XDR Memory interface bus width: 56bit out of 64bit (serial)

The RSX does not have a XDR memory interface, maybe you ment FLEXIO?

Edit: I see you say FLEX IO in the table just below
 
Cell:
65nm (212mm2 die size) technology

Are you sure about that? I thought it was 174.6mm2. 212mm2 might be the 65nm CELL with enhanced fp64 throughput, but I'm not sure about this...


RSX:
28 pixel shaders (4 redundant, 24 active)

Where does this come from? I've only seen specualtions about that, but nothing to suggest it's actually true. And if it's true there would be 28 texture units too.

That's just a more descriptive way of saying 6 Quads + 1 redundant Quad (4 PS pipes with a Texture Unit each) just like the 8th SPE in each CELL BE CPU is disabled to improve yields.
Other things have been mentioned by various devs on public forums (like the post transform vertex cache, etc...) or make sense if you think about it for a few seconds: like the 550 MHz clock for the Pixel Shader ALU's array knowing that nVIDIA has been using different clock domains across a GPU for quite a while now with different clock speed across the GPU: seeing that GDDR3's interface is at 650 MHz I'd guess that the ROP's are running at 500 MHz just like the rest of the chip beside the Pixel Shader ALU's which run at 550 MHz... not OFFICIALLY confirmed, but very believable. Devs have stated on public forums that RSX is certainly not weak at the Pixel Shading level...
 
That's just a more descriptive way of saying 6 Quads + 1 redundant Quad (4 PS pipes with a Texture Unit each) just like the 8th SPE in each CELL BE CPU is disabled to improve yields.
Other things have been mentioned by various devs on public forums (like the post transform vertex cache, etc...) or make sense if you think about it for a few seconds: like the 550 MHz clock for the Pixel Shader ALU's array knowing that nVIDIA has been using different clock domains across a GPU for quite a while now with different clock speed across the GPU: seeing that GDDR3's interface is at 650 MHz I'd guess that the ROP's are running at 500 MHz just like the rest of the chip beside the Pixel Shader ALU's which run at 550 MHz... not OFFICIALLY confirmed, but very believable. Devs have stated on public forums that RSX is certainly not weak at the Pixel Shading level...

Hmm that's a new one to me. I thought the settled perception was RSX was@500 mhz.

That would be pretty interesting news if true, but I wonder why they'd downclock only parts of the chip?
 
Hmm that's a new one to me. I thought the settled perception was RSX was@500 mhz.

That would be pretty interesting news if true, but I wonder why they'd downclock only parts of the chip?

They probably do that to get a better balance between power consumption and performance. Also it looks like most of the chip is running at 500 MHz with only the Pixel Shader ALU's running at 550 MHz if that page is accurate (nVIDIA specifies a clock speed rating for their GPU's but that does not mean all the portions of the chip run at that frequency).
 
I know that on the G70/G71 the vertex shader clock was not synced to the rest of the GPU. I think it was 470MHz/520MHz pixel/vertex clock for a standard 7900GT.
 
I know that on the G70/G71 the vertex shader clock was not synced to the rest of the GPU. I think it was 470MHz/520MHz pixel/vertex clock for a standard 7900GT.

Hmm seems they only ever ran the vertex shader clock faster anbd the shader clock slower with 7900GT not vice versa..seems unlikely they would reverse it.
 
Hmm that's a new one to me. I thought the settled perception was RSX was@500 mhz.

Well, it would still be RSX running at a core clock of 500MHz even if the pixel shaders were clocked at 550MHz. Personally I haven't heard the possibility before, but Pana's right that it'd be plausible. As for the redundant quad, well I've always believed that to be the case; Kutaragi seemed to allude to it in the past himself and for a die that size it would just make sense to set up an in-built defect-tolerance. It's pretty common for GPUs to work on that model anyway.

Of course this whole 550 domain thing might start a wave of rumors/excitement across the net claiming that RSX really does run at 550MHz.... but no, it still doesn't. People grab on to what they want to believe though. :)

And yes that 65nm Cell figure on that page is wrong; that's the die size for the HPC version of Cell, not the shrunk 'standard' version at 65nm.
 
Pretty neat compilation of information on the entire system.
So just what is the Secret of Monkey Island ?
:)
 
Thanks for all the feedback. I've corrected the 65nm die size and incorporated some related info. I think the size given is for IBM's own regular 65nm version, not the one by Sony, though.
 
Thanks for all the feedback. I've corrected the 65nm die size and incorporated some related info. I think the size given is for IBM's own regular 65nm version, not the one by Sony, though.

Speaking of die sizes, your 90nm size is too low; it's 235mm^2. You're showing the DD1 revision of Cell, which is not the one that actually went into production. And that's the correct 65nm die area, don't worry. At most, there might be a slight variance between chips fabbed at Fishkill and those at Nagasaki, but as the chip topography will be identical on the shrink, and the shrinks themselves are I/O limited, I'd be happy to guess they are mm for mm identical.
 
Last edited by a moderator:
I dug a little deeper about the DD1, but I think the original figure at http://www.edepot.com/playstation3.html is still correct.



IBM DD1 and DD2 Prototype Die Sizes

http://www.realworldtech.com/page.cfm?ArticleID=RWT072405191325



Measured Sony PS3 Cell Die Size (60gb 90nm)

http://techon.nikkeibp.co.jp/article/NEWS/20061124/124359/
(there are other sites, but this is one of them that actually measured)



More PSP info (includes flash too):
http://www.edepot.com/reviews_sony_psp.html


I did not know xbox360 had such a small flash.
 
Status
Not open for further replies.
Back
Top