Cost of Cell

[Carl B]

To get the disclaimers out of the way, this information comes from The Inquirer, so really there's no humanly way of knowing whether this is one of their BS stories or not. That said, it seems like it's not, but whatever...

Anyway acording to this article of theirs, presently IBM is yielding about 65 'good' Cell chips per wafer. Me and others have previously counted the die candidates on a wafer and reached 228 as the number of possible chips; thus at present if the 65 number is to be believed, the present yields are in the ~29% range. With an (Inquirer stated) total wafer processing cost of roughly $10,000 per wafer, this gives a pre-packaging cost of around ~$154 per chip.

Well, the Inquirer then goes on to talk about the price effect on PS3, blah blah blah...

But rather, I think they missed an important point of the 65 'good' Cell chips almost assuredly refering to perfect dies, whereas the PS3 will be utilizing chips with up to one SPE defective. So clearly if these numbers are true the cost to Sony for the PS3 will not be as high as that $154 per chip (pre-packaging of course), but rather some lower number.

Is there a means via which we could estimate how many more die candidates we might receive from a wafer at the present yields with single-SPE defective dies included? I know there is an equation to calculate roughly the estimated die defects on wafer and their spread, but I don't know it myself. (hint hint)

Though on a large die such as that for Cell it's hard to estimate how much of a gain in yield percentage we might be able to expect in the next several months pre-65nm, certainly some gain though should be expected - again aiding the situation of the chips cost.

Anyway thought this was interesting info; it of course depends wholly on the veracity of the Inq's 'information,' but it's the closest we've gotten thus far to a true snapshot of the Cell's present state of fabbing and cost.

Picture thrown in for dramatic effect.

 

[Guden Oden]

We must not forget tho that IBM is not going to be the sole supplier of cells either. Sony/Toshiba have their own fabs with production lines that are in the process of ramping up, or already have.

 

[Xenus]

I did a quick search and the most I could find for a 12 90nm SOI wafer was $3,000 and the cheapest was $600. Of course the tolerances were different and we don't know the tolerances needed for cell

 

[Carl B]

We must not forget tho that IBM is not going to be the sole supplier of cells either. Sony/Toshiba have their own fabs with production lines that are in the process of ramping up, or already have.

I completely agree, and Sony might in fact have better (or worse) yields than IBM as well. Certainly IBM will not constitute the sole source of chips for them, whatever the case. But all the outside variances aside, these numbers provide us with the first 'hard' numbers we've had to work with in a while, so if nothing else even if it only gives us insight into IBM's process, it'd be interesting to know some people's thoughts on the yield percentage when a defective SPE is taken into acoount, and further what some reasonable target yield percentage improvements might be for the near future.

I did a quick search and the most I could find for a 12 90nm SOI wafer was $3,000 and the cheapest was $600. Of course the tolerances were different and we don't know the tolerances needed for cell

Well that $10,000 cost wouldn't be for the wafer itself of course, but the costs associated with the entire fabbing process. Indeed the wafer is the cheapest part of the whole thing.

 

[Bobbler]

This is also talking just about IBM's fabbing costs and yields, not Sony's (for PS3). Ignoring the cost of the wafer, IBM's Cells that they have been using (for mercury systems and the like) are 8 SPE enabled so they are going to be more costly no matter (at least until they seemingly start using the "damaged" chips). Thats going to hurt yields this early on, big time.

 

[Carl B]

This is also talking just about IBM's fabbing costs and yields, not Sony's (for PS3). Ignoring the cost of the wafer, IBM's Cells that they have been using (for mercury systems and the like) are 8 SPE enabled so they are going to be more costly no matter (at least until they seemingly start using the "damaged" chips). Thats going to hurt yields this early on, big time.

Well, Sony will be sourcing a number of their Cell's from IBM though, and to be sure you can bet that any of these Cell chips with one SPE defective that are otherwise up to snuff aren't being tossed by IBM; rather they're probably being set aside for Sony. It's that estimated yield percentage with the defect tolerence worked in that I'm hoping we can make an educated guess at in this thread.

 

[The GameMaster]

It's quite pricey... no doubt about that, actually a tad more than I thought. Anyway... I could of sworn that Sony was going to handle the manufacturing of the Cell processors (as well as the RSX) in the PS3 in their own fabrication facilities. If IBM is fabricating these processors for Sony then that would make things even more expensive for Sony. I still don't know how expensive it would be for Sony to fabricate these processors themselves as it could very well be even higher...it is difficult to say for sure as we do not know how mature Sony's fabrication is versus IBM's (or TMSC's for that matter). Defect rates are also an issue with the fabrication of the Cell processor with as large of a die and as many transistors as that processor has.

Pretty interesting information...

 

[Carl B]

It's quite pricey... no doubt about that, actually a tad more than I thought. Anyway... I could of sworn that Sony was going to handle the manufacturing of the Cell processors (as well as the RSX) in the PS3 in their own fabrication facilities. If IBM is fabricating these processors for Sony then that would make things even more expensive for Sony. I still don't know how expensive it would be for Sony to fabricate these processors themselves as it could very well be even higher...it is difficult to say for sure as we do not know how mature Sony's fabrication is versus IBM's (or TMSC's for that matter). Defect rates are also an issue with the fabrication of the Cell processor with as large of a die and as many transistors as that processor has.

Pretty interesting information...

Sony has volume rights to a certain amount of IBM's 65nm East Fishkill output, an aspect of the deal they struck when Sony helped pay for the line expansion at East Fishkill. Link

In addition, IBM will be supplying Sony with Cell's on a normal basis from their 90nm lines as well; I would expect a lot of the Cell's in the launch PS3's to have originated at Fishkill, and I think an aspect of the STI agreement in general was that Sony has initial rights - should they choose to exercise them - to the Cell's produced by all STI partners leading up to PS3's launch. Sony's own 90nm lines are plenty mature - of that there is no doubt - the only question would be initial ramp-up yields as to my knowledge Nagasaki is not full-on fabbing Cell just yet. I'm not too worried about it. It may surprise you how high these initial fabbing costs are, but on the contrary, with the defect-tolerence the Cell's going into PS3 have built in, I'd expect the costs relative to these 'perfect' chips to be appreciably lower. An estimate of just how much lower hopefully being fruit this thread will bare at a later point.

 

[dodo3]

Hey xbdestroya, I don't know if you can answer this question, but do chips/hardware costs in manufacturing lower every other year?

Ex: When Sony started making the PlayStation 2 (original) did the cost to manufacture lower until the Slimline showed up?

Thanks in advance.

 

[Carl B]

Hey xbdestroya, I don't know if you can answer this question, but do chips/hardware costs in manufacturing lower every other year?

Ex: When Sony started making the PlayStation 2 (original) did the cost to manufacture lower until the Slimline showed up?

Thanks in advance.

Yeah, they do lower; a number of factors play a role.

But probably the most important one for achieving those cost savings are the process shrinks which occur throughout the component chips lives. Simply put, these shrinks should result in a rough halving of chip costs once fabrication on major new process nodes reach maturity; that is to say, every time a fab on a new node can approach half the die size, it's approaching half the cost of the original.

For the PS2 in particular, here's a slide showing the process nodes throughout it's life:

 

[dodo3]

Yeah, they do lower; a number of factors play a role.

But probably the most important one for achieving those cost savings are the process shrinks which occur throughout the component chips lives. Simply put, these shrinks should result in a rough halving of chip costs once fabrication on major new process nodes reach maturity; that is to say, every time a fab on a new node can approach half the die size, it's approaching half the cost of the original.

For the PS2 in particular, here's a slide showing the process nodes throughout it's life:

Oh nice, thanks a lot. So everytime the EE shrank, it halfed the cost?

 

[Shifty Geezer]

I doubt it halved the cost every shrink, as not every sjhrink was half the size, and I'm sure there's other costs involved on top of just wafer size. The diagram shows shrinks of 93%, 49% and 66% until the combined chip. I guess the final EE+GS at 1/6th the size is 1/6th the price. If the chipset cost Sony $200 back in the day, it'd now be about $30. Those are the sorts of savings MS isn't getting in XB. This is also where we see how the initial costs of PS3's chips are much less of a concern than long term price reductions. Initially there'll only be a few PS3s, too expensive for most buyers. Sony will bleed some money there to begin with. 3 years down the line when the hardware's mainstream is when Sony really need a cheap solution.

Regards the Cell results, this is something someone ought to know, but a Google didn't find me any straight predictions. Still at $10k a wafer, that's a minimum of $45 per chip. 50% yield is $90. That places the analysts $100 price-point in the realms of reasonable, though the high quality chips sold at a premium could offset the costs a bit.

 

[Tahir2]

To put things in perspective Intel was quoted to have said that each Pentium 4 chip wafer cost them $40 and evne with recent problems and die size increases it still costs approx $40.

 

[[maven]]

Is there a means via which we could estimate how many more die candidates we might receive from a wafer at the present yields with single-SPE defective dies included? I know there is an equation to calculate roughly the estimated die defects on wafer and their spread, but I don't know it myself. (hint hint)

You could use my DDS (Die Destruction Simulator) Python script...

 

[mesyn191]

To put things in perspective Intel was quoted to have said that each Pentium 4 chip wafer cost them $40 and evne with recent problems and die size increases it still costs approx $40.

Intel gets some of the best yields in the business from thier fabs though, don't know thier recent percentages but I do know it was announced several times that they were able to get 100% yields out of the P!!! on some runs and in the high 90's most other times...

Also this "comparison" ignores the fact that IBM expects to make a profit selling these chips to Sony so there is some sort of markup going on here, even in bulk shipments, no way IBM will sell these at cost of manufacture. I doubt Sony is paying over $150 per chip at the moment too, but I wouldn't be shocked to see them paying somewhere around $100 for each one right now. So long as yields increase quickly it doesn't really matter.

edits; spelling and such...

 

[Shifty Geezer]

Also this "comparison" ignores the fact that IBM expects to make a profit selling these chips to Sony so there is some sort of markup going on here, even in bulk shipments, no way IBM will sell these at cost of manufacture.

I don't think so. IBM aren't selling the chips to Sony. Sony put up the money to fund the research and build the fabs. If they then have to pay profits on top for the manufacturing, they got a really bad deal! Sony should be getting Cells at cost for PS3.

Likelihood is, I guess, IBM keep the working 1:8 Cells for their own customers, and give the 1:7 Cells to Sony. This builds some stock while Sony and Toshiba get their own plants working for the bulk of their requirements. So far it doesn't seem Toshiba have a high requirement for Cells but we don't know what they have planned.

 

[Carl B]

To put things in perspective Intel was quoted to have said that each Pentium 4 chip wafer cost them $40 and evne with recent problems and die size increases it still costs approx $40.

Like was already mentioned, Intel's kind of a special case. Indeed $40 is their supposed 'per die' cost, but Intel's got some of the best fabbing operations in the industry and remember also that the P4 is around half the size of Cell to begin with.

@Maven: Thanks for the script. Unfortunately I can't run it myself, but maybe later I'll try to distill the equation from it and see if I can't use the stated yield percentage of 65:228 to reverse engineer the defect rate and get us to what the yields for 7 functional SPE's should/would be.

 

[hongcho]

The funny thing is... Recently I was talking to one of my college friends who works at IBM Research in Austin, TX (his main research theme is on optimizing circuit routing software). He said IBM does NOT (I assumed he meant "usually") care about the yield that much (as compared to Intel): even if only one come out good in a wafer, IBM thinks it's a success. I don't think he was referring to IBM's concole chips, but for those server chips where the cost structure is completely different, IBM does not care too much about the wafer yields.

Hong.

 

[mesyn191]

I don't think so. IBM aren't selling the chips to Sony. Sony put up the money to fund the research and build the fabs. If they then have to pay profits on top for the manufacturing, they got a really bad deal! Sony should be getting Cells at cost for PS3.

IIRC they only put up some of the money to research Cell and to build the fab (added up to less than a half a billion didn't it?), most of it was out of pocket costs for IBM...

Helping to pay for fab construction & research != Cell for no additional cost

Likelihood is, I guess, IBM keep the working 1:8 Cells for their own customers, and give the 1:7 Cells to Sony. This builds some stock while Sony and Toshiba get their own plants working for the bulk of their requirements. So far it doesn't seem Toshiba have a high requirement for Cells but we don't know what they have planned.

I doubt they're keeping more than a handful, almost all of em' must be going for PS3 production ramp ATM.

 

[London Geezer]

10k for one big glass disc... wouldn't wanna drop that!!!