Does DD2 signify a problem with Cell?

nelg

Veteran
According to rumours (or maybe facts) the the PPE in the DD2 cell is twice the size of the PPE in the DD1 cell and is the only change from one to the other. This does not seem like a minor revision at all. Does this signify an attempt to rectify a glaring weakness? Does anyone know the why there was such a major revision?
 
Theory is because they were trying to flesh out the Power core to garner Apple support - but whatever the case, they made it bigger/better.

By the way, they're on DD3.1 at least, so there may be other quasi-significant changes that have taken place as well (though none as big as DD1--<DD2 I imagine.)

Anyway I don't see it as a problem at all; rather an expansion of the PPE at the cost of die-size. Since we can expect the PPE will be relied upon heavily in the early stages of Cell in terms of programming, it's probably a good move until developers/programmers become better acquainted with the SPE's. As time goes on and there are process shrinks, the 'doubling' of the PPE will become an increasingly trivial action in terms on it's effect on die area and yields.
 
Last edited by a moderator:
xbdestroya said:
Theory is because they were trying to flesh out the Power core to garner Apple support - but whatever the case, they made it bigger/better.

By the way, they're on DD3.1 at least, so there may be other quasi-significant changes that have taken place as well (though none as big as DD1--<DD2 I imagine.)

DD3 was apparently just tweaking to improve yields.
 
Titanio said:
DD3 was apparently just tweaking to improve yields.

Well that's what they say. ;)

But I mean I think something like DD3-->D3.1 - that sounds like tweaking for yields. Not saying DD3 is anything exciting, but I'm just surprised it wouldn't have been given a DD2.x sub-designator if it were *just* for yield purposes.
 
xbdestroya said:
Theory is because they were trying to flesh out the Power core to garner Apple support - but whatever the case, they made it bigger/better.

I did not hear of this before. Considering the time frame how plausible is the rumour?
 
nelg said:
I did not hear of this before. Considering the time frame how plausible is the rumour?

No idea. I think honestly the second part of my post is the more plausible - simply seeking greater single-threaded performance with early development efforts in mind.

But if you want to explore the rumor, this is probably the best place to read about that angle:

http://www.realworldtech.com/page.cfm?ArticleID=RWT072405191325
 
version said:
DD1 - 1 instruction/cycle
DD2 - 2 instructions/cycle
DD3 - 4 instructions/cycle

Which raises the my original question. This would have been known before as per its design. So after getting back silicon for testing did the instruction/per cycle throughput in the PPE prove to be such a limiting factor that a redesign was necessary?
 
I don't know how you could qualify it as either necessary or desirable. the fact a chip improves doesn't mean it was bad to begin with and neded to be good, but could mean it was good and is now better. Or, it could mean is wasn't a very good idea and they had to revise it. But we've no way of knowing, short of getting our hands on DD1 and DD2 and DD3 revisions and benchmarking them to see if DD1 really was a poor performance part.
 
According to rumours (or maybe facts) the the PPE in the DD2 cell is twice the size of the PPE in the DD1 cell and is the only change from one to the other. This does not seem like a minor revision at all. Does this signify an attempt to rectify a glaring weakness? Does anyone know the why there was such a major revision?

I assume there was a good reason behind it but I've no idea what it is. I assume it's a better core.

The Cell is a completely new design and wasn't just built and churned out. They have made big changes in order to get things right.
The Cell described in the patent is different from the Cell finally released, the SPEs local stores were doubled in size twice and the SPEs themselves underwent a major redesign at some stage.

The Cell in the PS3 probably won't change significantly after it starts production but Sony has already mentioned future revisions which can go elsewhere.

DD1 - 1 instruction/cycle
DD2 - 2 instructions/cycle
DD3 - 4 instructions/cycle

Any truth in that? 4 instructions per cycle implies the PPE either got replaced again or at least underwent major changes.
 
I'm also thinking that even in the long run, PPE utilization will be pretty heavy because even after developers get comfortable with the SPEs, there will be more data moving in and out, and management of all that is probably best handled at the PPE. If they did widen the ways of SMT, that seems fairly appropriate, especially for hiding more latency.

I'm still waiting to see some die photos that show that something actually happened. DD2 photos are there as are some number of associated specs and studies, but I've yet to see anything remotely like that on DD3.
 
Shifty Geezer said:
I don't know how you could qualify it as either necessary or desirable. the fact a chip improves doesn't mean it was bad to begin with and neded to be good, but could mean it was good and is now better. Or, it could mean is wasn't a very good idea and they had to revise it. But we've no way of knowing, short of getting our hands on DD1 and DD2 and DD3 revisions and benchmarking them to see if DD1 really was a poor performance part.

Of course this is really a call to speculate since the only ones who would know are under NDA. That being said, it seems like such a major revision that there was some mistaken estimations. Or perhaps, like xbdestroya mentioned, it is to facilitate first generation games until the transition to the new programming model is more mature and the SPEs can be utilized more independently.
 
ADEX said:
Any truth in that? 4 instructions per cycle implies the PPE either got replaced again or at least underwent major changes.

Please read versions posting history before taking anything he posts seriously.....
 
Assuming that the data structure is designed and prep'ed properly, can the PPE delegate synchronization/management/orchestration work to a SPE and has that SPE deal with the rest of the SPEs ?
 
MrWibble said:
I can't wait for DD7! 128 instructions/cycle!
Now you're just being ridiculous. The progression is such that for the nth revision of PPE (DDn), number of instructions per second is 2^(n-1). We can only really expect to get 64 instructions/cycle from DD7, realistically speaking.

Edit : I'd better add a smiley as without someone's sure to think I'm being serious :p
 
patsu said:
Assuming that the data structure is designed and prep'ed properly, can the PPE delegate synchronization/management/orchestration work to a SPE and has that SPE deal with the rest of the SPEs ?

SPEs can self manage, yes.

PPE would require only minimal involvement at boot-time if that's what you wanted to do.
 
Shifty Geezer said:
Now you're just being ridiculous. The progression is such that for the nth revision of PPE (DDn), number of instructions per second is 2^(n-1). We can only really expect to get 64 instructions/cycle from DD7, realistically speaking.

Edit : I'd better add a smiley as without someone's sure to think I'm being serious :p

Well I was of course assuming an extra DD4.5 version to throw the progression off! :)

But hey, I'd settle for 64...
 
MrWibble said:
SPEs can self manage, yes.

PPE would require only minimal involvement at boot-time if that's what you wanted to do.

Thanks for the info. Yes, as an incremental step to see how far PPE-centric development approach can go. But I'd think like others have mentioned, it's probably better to redesign the whole thing for Cell.
 
Last edited by a moderator:
patsu said:
Thanks for the info. Yes, as an incremental step to see how far this development approach can go. But I'd think like others have mentioned, it's probably better to redesign the whole thing for Cell.

I'm not sure what you mean - as a rule we find having the SPUs self manage is much more "cell friendly" than having the PPU get involved.

I consider the PPU to be more of a vestigial tail than a vital organ.
 
Back
Top