Interesting interview with IBM (Cell, backwards compatability...)

[SugarCoat]

50% of the die area of Cell is made of SPEs, right ? So if a chip has a single defect, there's a 50% chance it can still be used for the PS3 (and if it has several defects, there is a (much smaller) chance all the defects end in the same SPE area).

depends where the problem is. If its in the PPU the chip is completely junk. 10-20% is also quite a leap, so i would say its on the worse side, which isnt exactly shocking. The yields of Cell with 7 wont be much better. I dont know how anyone can put a happy face on this. They need millions of them as opposed to graphics cores which people are using for comparison, which in the high end dont generally sell more then a say, ten or twenty thousand (on a great product). Theres no comparison since the Cell has to meet such a huge quota, losses on bad yeilds will be MUCH greater. Plus Sony isnt making a profit off the processor within the PS3 unlike Nvidia and ATI.

The cell is in yield hell?

Who soaks up the costs of failures? Sony or IBM?

I do appreciate his honesty and not PR talk.

Plain and simple, this is bad news for PS3. If these are Sony's orders Sony is taking the loss, not IBM. Sony pays for the silicone whether its working or not.

 

[ZiFF]

They will be moving to 65nm in the spring?

 

[Arwin]

Plain and simple, this is bad news for PS3. If these are Sony's orders Sony is taking the loss, not IBM. Sony pays for the silicone whether its working or not.

Plain and simple, this is NO news for PS3. Nobody would have assumed different yields, IBM isn't saying that they are lower than expected; they are exactly what is expected, and the Cell has been in production for quite a while.

IF there are going to be problems for the PS3, I very much doubt the Cell would be one of them.

 

[expletive]

IF there are going to be problems for the PS3, I very much doubt the Cell would be one of them.

Why? If the PS3 experiences manufacturing problems, I would expect this or the BR drive to be the MOST probable culprits. Which is to say, the most cutting-edge and untested (in terms of large scale manufacturing) technology.

 

[TheChefO]

Why? If the PS3 experiences manufacturing problems, I would expect this or the BR drive to be the MOST probable culprits. Which is to say, the most cutting-edge and untested (in terms of large scale manufacturing) technology.

And Ram. Xdr is pretty cutting edge as well and could have manufacturing issues too.

Overall I expect them to come through though.

 

[Arwin]

Why? If the PS3 experiences manufacturing problems, I would expect this or the BR drive to be the MOST probable culprits. Which is to say, the most cutting-edge and untested (in terms of large scale manufacturing) technology.

I think it is a mistake to confuse the manufacturing process of the Cell chip with the design of the Cell chip. The design of the Cell chip is cutting-edge, but the manufacturing process with which it is produced is much less so. Mind you, I'm not an expert on these things, but that's how I understand it.

 

[Platon]

I think it is a mistake to confuse the manufacturing process of the Cell chip with the design of the Cell chip. The design of the Cell chip is cutting-edge, but the manufacturing process with which it is produced is much less so. Mind you, I'm not an expert on these things, but that's how I understand it.

That is true, the manufacturing process is the same for more or less all the chips now, that does not mean that the design/complexity of the chip does not affect its manufacturing, not to mention that the cell chip is quite huge and the bigger the worst yields they have...

 

[Jawed]

For example, in the shift from G70 to G71, the number of transistors dropped, and in some reviews I have seen this partially attributed to the removal of redundant logic within shaders.

Wrong kind of "redundant". In this case redundancy arises because the 90nm process allows the pipeline to be clocked higher more easily, thus less stages in the pipeline are required to achieve target clocks.

Jawed

 

[Titanio]

Thanks for the insight.

A more general logic redundancy approach may make sense with complicated chips though, I'd be kind of surprised if it's not been done before. That's just my lay-man's point of view..it's probably more complicated than adding in a few extra transistors

 

[Jawed]

I still hold that in context, he was referring to yields from 4- upto 8-SPE Cells. The logic redundancy he was referring to was nothing more than the sacrifice of an SPE. Since PS3 was architected on this very basis, I really can't see why anyone here is arguing that the sacrificial SPE doubles yield and is precisely what he was alluding to.

Jawed

 

[Rolf N]

I still hold that in context, he was referring to yields from 4- upto 8-SPE Cells. The logic redundancy he was referring to was nothing more than the sacrifice of an SPE. Since PS3 was architected on this very basis, I really can't see why anyone here is arguing that the sacrificial SPE doubles yield and is precisely what he was alluding to.

Jawed

Because he also said that this type of logic redundancy hadn't been done before. I don't know about you, but I find "dud quads" in a graphics chip very comparable to a borked SPE, and I'm also pretty sure the guy had heard about ATI and NVIDIA doing that kind of thing. Still he made that claim.

That's why.

 

[PiNkY]

Plain and simple, this is NO news for PS3. Nobody would have assumed different yields, IBM isn't saying that they are lower than expected; they are exactly what is expected, and the Cell has been in production for quite a while.

IF there are going to be problems for the PS3, I very much doubt the Cell would be one of them.

You'll likely never see any accurate yield figures in Internet forums as sc companies typically consider these confidential. It is however not unusual to have yields below 50% during ramp-up. However for mature products they are typically in the 90% range. At least that is what I remember from University talks. These numbers in the interview stroke me as rather low, for a mass-market product being fabbed for close to a year now (not to mention that he disclosed them, but then, he wanted to drive home a point about redundancy). I would say, that if there are any hurdles with the PS3's production, they will most likely stem from its most complex/new parts cpu, br-drive & gpu (in that order).

 

[Jawed]

Because he also said that this type of logic redundancy hadn't been done before. I don't know about you, but I find "dud quads" in a graphics chip very comparable to a borked SPE, and I'm also pretty sure the guy had heard about ATI and NVIDIA doing that kind of thing. Still he made that claim.

That's why.

What about other CPUs? And not counting memory on those CPUs? Why should he be including GPUs in his generalisation? - he's a CPU guy.

Considering how much trouble IBM has had making fast Power CPUs, why is the 8-SPE yield surprising? Why do people think it won't get hugely better, and with 65nm better again?

Jawed

 

[RobertR1]

depends where the problem is. If its in the PPU the chip is completely junk. 10-20% is also quite a leap, so i would say its on the worse side, which isnt exactly shocking. The yields of Cell with 7 wont be much better. I dont know how anyone can put a happy face on this. They need millions of them as opposed to graphics cores which people are using for comparison, which in the high end dont generally sell more then a say, ten or twenty thousand (on a great product). Theres no comparison since the Cell has to meet such a huge quota, losses on bad yeilds will be MUCH greater. Plus Sony isnt making a profit off the processor within the PS3 unlike Nvidia and ATI.



Plain and simple, this is bad news for PS3. If these are Sony's orders Sony is taking the loss, not IBM. Sony pays for the silicone whether its working or not.

Thanks Sugarcoat for the exact answer.

In terms of scaling, let say they have even a 25% yield return. For 3 million working chips, they'll have to produce 12million and figure out what to do with other 9million. None of this chips can obviously be used for a lesser sku's. His comment about "not throwing away the 6spe chips" alludes to me that Sony has no use for them.

I wonder if the Cell is customized in the PS3 to the point that failed chips cannot be used for other applicatons?

 

[arjan de lumens]

What about other CPUs? And not counting memory on those CPUs? Why should he be including GPUs in his generalisation? - he's a CPU guy.

Dual-core processors with 1 good + 1 bad core are often sold as single-core processors; in particular, AMD is known for doing this with the "Manchester" revision of Athlon64.

 

[Jawed]

Dual-core processors with 1 good + 1 bad core are often sold as single-core processors; in particular, AMD is known for doing this with the "Manchester" revision of Athlon64.

http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2484

That was solving cache failure, not core failure. I specifically excluded memory.

Jawed

 

[arjan de lumens]

http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2484

That was solving cache failure, not core failure. I specifically excluded memory.

Jawed

And so did I. There are plenty of reports in the wild about "manchester" cores that are sold with not only reduced cache but also a core disabled. Just do a google search for e.g. "AMD manchester 3500+".

EDIT:
and here is something that appears to be a confirmation directly from an employee of AMD itself:
http://forums.amd.com/index.php?showtopic=71281#

 

[Jawed]

http://forums.amd.com/index.php?showtopic=62563

Better tell the IBM guy then

Jawed

 

[Rolf N]

What about other CPUs? And not counting memory on those CPUs? Why should he be including GPUs in his generalisation? - he's a CPU guy.

It's possible that he doesn't consider memory redundancy to be the same thing as logic redundancy. And yes, it's also a possibility that he ignores GPUs.

This isn't clear enough however to proclaim that the discussion what exactly he means when he talks "logic redundancy" should stop IMO. It's not settled.
When I read your "I really can't see why anyone here is arguing that the sacrificial SPE doubles yield and is precisely what he was alluding to" I thought that's what you wanted to say, but going back and rereading it now it seems that you didn't quite say that, and I wasn't particularly clear enough on what I wanted from you either.

What I really wanted to talk about was their "logic redundancy" which IMO is not necessarily a reference to a disabled SPE in light of the "we're first" comments. If he thought they were first with that, he's not well informed. If he was aware that they weren't first and said it anyway, well, let's just say that's not IBM's current style AFAIK.

Considering how much trouble IBM has had making fast Power CPUs, why is the 8-SPE yield surprising? Why do people think it won't get hugely better, and with 65nm better again?

Jawed

Me?
I'm just going for the more optimistic one of the two possible interpretations of the IBM-guy statements. I don't think that's unreasonable ...

 

[Jawed]

Taking a different tack, entirely: IBM's (or STI to be fair) pretty proud of Cell and there's lots of nice technical stuff out there about how and why it's been put together.

If there's a kind of logic redundancy (expressly for yields) that's unique and new to Cell, don't you think there'd be a paper about it?

All we have is the heavily documented 7-SPE PS3-Cell. Was it always planned to be 7-SPEs (i.e. when they settled on the 8-SPE design, was it cognisant of early-life yield by electing to use 7 for PS3)? It's too long since I've read PS3-specific stuff on this topic for me to remember

I dunno how much you can tell from a 1930x3000 pixel image of the die:

http://www-03.ibm.com/chips/photolibrary/photo10.nsf/WebViewNumber/ED994790FAECFD6900256FEA0062126B

but perhaps someone should go fishing for logic-level redundancy that's at a level lower than an entire SPE? Comparing blocks such as floating point or instruction decode etc. against near relatives from other CPUs in the PowerPC line and perhaps coming back and saying "look the Cell Vec4 pipeline has redundancy".

---

Also, with the AMD dual-cores->single-core, it seems to me that AMD is a bit embarrassed about it - otherwise why did enthusiasts have to spend so much effort getting AMD to admit to it?

---

Which was the first GPU to bin by dead pipes/quads?

---

NVidia has a patent on the subject, which could go back to Geforce 6800:

http://www.beyond3d.com/forum/showthread.php?t=25332

And here's some of ATI's:

http://www.beyond3d.com/forum/showthread.php?t=29119

which seem to be more fine-grained than an entire quad being switched off. Sadly the links to the USPTO in that thread now point at completely different patents!!!!! ARGH.

Jawed