I believe RSX will have a qaud disabled

[Bill]

It makes perfect sense:

Yields. Obviously PS3 will need to be mass produced on an extreme scale. Xenos (apparantly) and Cell both have redundancy built in. Why not RSX?

Aggressive clock target (granted on 90 nm)

Lines up with certain "odd statements". IE, Nvidia saying a more powerful GPU will be out at PS3 launch. Even an "Ultra" 7800 would be hard pressed to be higher than 550 mhz. However, if the RSX is 20 pipes, problem with statement solved.

Also supposedly on PSM DVD a Nvid engineer stated RSX<7800GTX by a little. Now I have't seen this so it's rumour, but if true, very powerful statement.

btw, a 550 20 pipe card about eqaul to a 26 pipe 430 mhz card. So RSX still>stock 7800 GTX.

Anyway, I just want credit when it happens!

 

[Titanio]

I believe the PSM thing was debunked.

It's very possible, Kutaragi has hinted at as much, but we don't really know where they're starting from with RSX.

I'm not really sure how it'd fit in with the numbers NVidia gave, but there was a thread on this very recently that you might want to have a look at:

http://www.beyond3d.com/forum/showthread.php?t=24211

 

[Bill]

The statement on paper was debunked (somewhat). Besides, a qaud disabled RSX still>stock clock GTX so their's no issue either way. However, this is supposedly on the DVD. I admit I haven't seen it.

 

[Carl B]

Well and besides that thread, the notion has come up several times. I myself have stated before that a quad might go disabled, and I'm hardly alone - so don't get too pleased with yourself just yet Bill.

Anyway as for NVidia having a 'more powerful' GPU come RSX's launch, I think the G72 - which should be at 90nm (speed, pipe bump?) - along with any number of other upcoming chips, should in theory qualify. RSX is supposed to launch spring next year, so by that time we'll have R580, NVidia's counters, and the next next-gen of graphics just months away.

 

[Acert93]

Chip-on-Chip, ignoring everything else: RSX > G70 (7800GTX).

Obviously some questions about memory; but on the other hand RSX will be in a situation with a much more fixed resolution (thus more optimization) and will be in a position, as being in a closed box, to see significantly much better performance than the G70 ever will.

So even if the RSX chip was less powerful (but it is not) I would expect it to be more powerful in the end results--and that is all that matters.

 

[Shifty Geezer]

Xenos (apparantly) and Cell both have redundancy built in. Why not RSX?

Where'd you here about Xenos having redundancy? That's the first I've heard. Being spread over two seperate dies seems to be the only approach at improving yields ATi have taken. I remember hearing nVidia (I think) talk of adding redundancy to RSX for yields, but not Xenos.

 

[Acert93]

Where'd you here about Xenos having redundancy?

It has been stated that there are both redundant ALUs in the logic AND redundancy built into the eDRAM as well. I believe this has been stated by a number of sources (I am pretty sure one is Dave). Anyhow, mentioned on the forums here (good luck finding it with the search feature... threads are HUGE and long so finding it may not be easy).

That's the first I've heard. Being spread over two seperate dies seems to be the only approach at improving yields ATi have taken. I remember hearing nVidia (I think) talk of adding redundancy to RSX for yields, but not Xenos.

You need to follow Xenos more

 

[Jawed]

If anything I'd hope that the relatively low speed and reasonable size of RSX (i.e. 550MHz and 300m transistors) on 90nm should make for something that has a good yield.

Sony's 90nm process has been in production for a solid year or so now. I expect Sony and NVidia have been focussed on that for a good 18 months.

NVidia should have had its first working G70s about 18 months ago from TSMC.

Overall I think it's more a question of getting RSX working (since it's not TSMC making it) than a worry about yields.

G70 is supposedly yielding very well. Once RSX is working, it should too, since it'll be less die space.

Jawed

 

[Jawed]

I think Xenos prolly does have four unified shader arrays, with one given-up for yield:

Jawed

 

[Acert93]

I want my 4th array (droooool) Can we use RivaTuner to enable it hehehe

 

[Love_In_Rio]

Those are 16 plus ALUs!!! I don´t think so. I believe that there may be some extra ALUs per shader array, being 2 or 3 disabled.

 

[Shifty Geezer]

Oooooh, photographic evidence of disabled ALUs! News to me and my ignorance

 

[Acert93]

Hmmm each of those ALUs only make up 10-15% of the chip? What are the odds they would be defective? I guess if the rest of the chip was designed to only handle 3 arrays enabling the extra one would be useless, but it has me wondering... they really are not that big!

 

[Jawed]

I estimate that the 8 areas I've coloured amount to about 32% of the die.

So each colour is about 8%.

My eyes always glaze over when people start talking the statistics of die yields, so I'll leave that to someone else...

Jawed