Predict: The Next Generation Console Tech

[TheChefO]

Right, so if MS did occupy 5% of their capacity, that would mean 300,000 chips per month @ > $50 per chip (at the minimum and not including costs for QA).

Actually a bit more expensive than that as listed above.

So at the 5% rate they would need 3 months to meet the million mark by the end of the year ... or just live with shortages ... or just launch in limited territories (US).

Or TSMC manufactures more than 60k/mo or gets better than 50% yields or MS requires more than 5% ...

 

[TheChefO]

That's... pretty expensive.

http://www.afterdawn.com/news/article.cfm/2008/12/31/isuppli_releases_bom_report_on_playstation_3_costs

isupply listed BOM for Cell and RSX at $46.46 + $58.01 respectively.... this is from 2008.

BOM prior is estimated at $59.47 + $75.41

 

[AlNom]

Actually a bit more expensive than that as listed above.

So at the 5% rate they would need 3 months to meet the million mark by the end of the year ... or just live with shortages ... or just launch in limited territories (US).

Or TSMC manufactures more than 60k/mo or gets better than 50% yields or MS requires more than 5% ...

Look, I don't know what the point is you're trying to get at with all this random guessing, but it is quite possibly very useless for the purpose of the thread because your hopes and dreams are founded in suppositions with zero real information. You're just making up calculations and going through the motions for the sake of arguing when there's no real substance. It's numerical assumption upon further assumptions that really don't get us anywhere real. The calculations you want regarding yield and chips per wafer etc are all founded on thin air.

Why are we assuming 5% of TSMC's capacity? Just because it means you'll hit 1 million units for launch? Why are we going with any specific number of wafers per month when there's nothing much to go on for how they will ramp up? Why specifically "10-25%" cost increase based on vague comments that don't even mention numbers relating to cost?

There are so many assumptions being made here in these calculations of yours that I'm not seeing the relevance to trying to predict something meaningful. (What if I picked a different random number)

 

[TheChefO]

... It's numerical assumption upon further assumptions that really don't get us anywhere real.

Actually the wafer numbers are real, the production is real, the chip sizes are real (based on gpu+ cpu sizes this gen at launch) the wafer cost estimates are in the ballpark.


The point I'm getting at is that it isn't impossible, isn't unreasonable, and isn't unthinkable that 2012 on 28nm is a very real possibility.


If you have numbers to counter those that I proposed, be my guest, but I've done my research and based on the info that is available, these are reasonable assumptions.



Let's narrow it down to a range:

20k is current production at TSMC.

Let's assume they can't ramp production at all ... 20k/mo wafers

4m/mo 300mm^2 chips @ 50% yield (pretty reasonable I'd say) = 2m/mo capacity - draw from this what you will on MS requirements.

100k/mo wafers = 10m/mo 300mm^2 chips @ 50% yield is their fab capacity.

So 2-10m units per month is a reasonable production range to expect in 2012.



As for costs:

http://semiaccurate.com/2011/09/08/exclusive-tsmc-raises-prices-on-amd-and-nvidia/

In the end, we hear that prices for 28nm products just went up by somewhere between 15-25%, lets call it low 20′s. If 28nm wafers cost about $5000, this is around a $1000+ bump.

So my target of $7,500/wafer should be rather generous and the upper limit of what to expect.

 

[AlphaWolf]

How many of those 20k wafers they are producing right now, have 300mm dies on them? And what is their yield?

 

[Ninjaprime]

http://www.afterdawn.com/news/article.cfm/2008/12/31/isuppli_releases_bom_report_on_playstation_3_costs

isupply listed BOM for Cell and RSX at $46.46 + $58.01 respectively.... this is from 2008.

BOM prior is estimated at $59.47 + $75.41

PS3 also launched at $500 and $600. It was also "sold at an estimated loss of $306 or $241 depending on model." Its sales also sucked until the price was dropped. I would bet 10 to 1 that no one is launching a $500 or $600 console this time around. Maybe not even $400.

 

[TheChefO]

How many of those 20k wafers they are producing right now, have 300mm dies on them? And what is their yield?

Dunno ... do you?

60-80% yields are reasonable based on what I've seen (depending on size, smaller = better yields), but I lowered it to 50% so as to account for the new process and relatively large expected GPU.

 

[TheChefO]

PS3 also launched at $500 and $600. It was also "sold at an estimated loss of $306 or $241 depending on model." Its sales also sucked until the price was dropped. I would bet 10 to 1 that no one is launching a $500 or $600 console this time around. Maybe not even $400.

PS3 had BRD and HDD which accounted for much of the initial cost.

Die size budget of xb360 was roughly the same as ps3.

 

[Ninjaprime]

PS3 had BRD and HDD which accounted for much of the initial cost.

Die size budget of xb360 was roughly the same as ps3.

You sure about that last part? Maybe I'm way off, but IIRC Cell was quite large, 220mm^2 or so, where as Xenon was 170mm^2 I think? RSX was "over 300 million" transistors, relative to Xenos's 232 mil on the same process, so it was most certainly smaller. Unless you're counting the eDRAM die, which I'm not sure you can count as die area, its probably cheaper than a logic die. Even counting the eDRAM I think you end up with less on the 360 side.

Edit: Cell was 230mm^2 and RSX was 258mm2 from a quick search, and Xenon is 176mm^2 and Xenos is 182mm^2. eDRAM die is 80mm^2.

 

[Ninjaprime]

Yeah, I seem to recall rumors of sub 40% yields on the last new process, and I think those were binned yields. Unless this one is magically better, when the trend is worse yields rather than better.

Edit: ^ Responding to alphawolfs comment, which went poof as I typed this I guess?

 

[hoho]

60-80% yields are reasonable based on what I've seen

Do tell what you've see. In the past 6-something years I've never seen any chip manufacturer release absolutely any concrete data on the yields.

 

[DuckThor Evil]

I don't think we are necessarily disagreeing, but what homer and I were taking issue with was pushing a 560 style card down into the mid range or as some were calling it mainstream market segment. It just isn't that sort of chip. I can see where hardcore PC gamers would see it as such, but that just isn't how these products are slotted.

That is all we were saying--I am not sure when roughly 80% of the performance (give or take) of your top single GPU solution becomes "mainstream" or "mid range" I think a fresh look of what the GPU landscape really looks like is in order. But I agree, a 560 is not in the same class as a 580, but homer was right to insist that there were more segments and categories in the market than others were allowing.

A lot of these "mainstream" "performance" etc. are just a play on words like lowest details in Crysis 2 is high "mainstream" is more and more low end these days it just sounds better... for a while.

Clearly if you look at performance, it's a fact that GTX 560 is way above 50% of GTX 580. It's die size is about 70%, but a very important fact is that the price of GTX 560 Ti is pretty much an exact 50% of 580.

499$ GTX 580
349$ GTX 570
249$ GTX 560 Ti
199$ GTX 560
129-149$ GTX 550 Ti

That is the range we need to look. Anything under is "out of range" for a discrete products at the lower end and 2 chip cards at the top are very marginal. Price is the main differentiator between products in their placements. The competition in the marketplace has made it so that in the 200-250$ segment the product features/performance is above what the price bracket alone suggests and the additional investment doesn't net you nearly as much as it used to couple years ago, but the price gap still is a very meaningful metric when we try to slot these things in to their places.

I'm proud of myself that I didn't bring out car analogies

 

[TheChefO]

You sure about that last part? Maybe I'm way off, but IIRC Cell was quite large, 220mm^2 or so, where as Xenon was 170mm^2 I think? RSX was "over 300 million" transistors, relative to Xenos's 232 mil on the same process, so it was most certainly smaller. Unless you're counting the eDRAM die, which I'm not sure you can count as die area, its probably cheaper than a logic die. Even counting the eDRAM I think you end up with less on the 360 side.

Edit: Cell was 230mm^2 and RSX was 258mm2 from a quick search, and Xenon is 176mm^2 and Xenos is 182mm^2. eDRAM die is 80mm^2.

I remember both xb360 and ps3 die areas being roughly the same (xgpu + xcpu + EDRAM = Cell+RSX).

Close enough for future projections.

 

[TheChefO]

Do tell what you've see. In the past 6-something years I've never seen any chip manufacturer release absolutely any concrete data on the yields.

http://www.semiwiki.com/forum/content/705-tsmc-28nm-20nm-silicon-update.html

I asked people who actually have 28nm silicon how it is going. Unanimously it was, “TSMC 28nm yield is very good!” Altera and Xilinx are already shipping 28nm parts .

http://www.xbitlabs.com/news/other/display/20091109232633_TSMC_Denies_Huge_40nm_Yield_Drop.html

While TSMC confirmed it did have certain issues with 40nm process technology, the world’s largest contract maker of semiconductors stressed that those issues were “logistical”. Thus, TSMC wants to stress that 40nm yields now are about 60%.

 

[Rangers]

Charlie seems like he no idea what he's talking about.

 

[TheChefO]

Charlie seems like he no idea what he's talking about.

If Charlie is claiming 28nm is impossible for 2012, I'm inclined to agree.

 

[hoho]

"40nm yields now are about 60%"

With what kind of die size? Is it a ram chip with mountains of redundancy that you can shoot to swiss cheese and still make it work or stuff that you'll throw away even if you get a single defect?

Also that 60% of whatever was at least a year after they started producing 40nm chips and even longer after 45nm.

 

[AlNom]

If Charlie is claiming 28nm is impossible for 2012, I'm inclined to agree.

No one is saying it's impossible.

 

[TheChefO]

Yeah, I seem to recall rumors of sub 40% yields on the last new process, and I think those were binned yields. Unless this one is magically better...

Indeed it is.

"The number of customer 28nm production tape outs has more than doubled as compared with that of 40nm. At 28nm, there are currently more than 80 customer product tape-outs. The TSMC 28nm process has surpassed the previous generation's production ramps and product yield at the same point in time due to closer and earlier collaboration with customers."

http://www.techspot.com/news/46013-tsmc-28nm-hits-mass-production-gtx-560-ti-refresh-coming.html

"I think 28nm is going to be a fabulous node," Moshe Gavrielov, CEO of Xilinx, tells me "a lot of learning went into 28nm to avoid the problems of 40nm. Four flavours, aimed at various markets, will give it tremendous longevity."

"The yields on 28nm are very good," says Gavrielov, "much better than they were at 40nm."

TSMC is ramping its 28nm process three times as fast as the company ramped the 40nm node.

"I've been there for the launch of the 65nm and 40nm nodes and have never seen five customers ramp a new technology as they're doing with 28nm - it's got three times the momentum of 40nm," Maria Marced, President of TSMC Europe, tells me.

"The number of tape-outs on 28nm compared to 40nm has been triple at the same stage of production."

http://www.electronicsweekly.com/blogs/david-manners-semiconductor-blog/2011/11/will-28nm-be-a-good-node.html



I dunno ... sounds like things are going smooth to me ...

 

[hoho]

"The TSMC 28nm process has surpassed the previous generation's production ramps and product yield at the same point in time due to closer and earlier collaboration with customers"

How long time has passed since first 28nm things rolled out from their product lines? Surprassing previous process in first few months isn't anything remarkable.