Maynard Handley (name99.delete@this.name99.org) on September 29, 2013 4:50 pm wrote:
> Maynard Handley (name99.delete@this.name99.org) on September 29, 2013 3:08 pm wrote:
> > mas (a.delete@this.b.com) on September 29, 2013 11:12 am wrote:
> > > bakaneko (nyan.delete@this.hyan.wan) on September 29, 2013 9:48 am wrote:
> > > > mas (a.delete@this.b.com) on September 29, 2013 9:31 am wrote:
> > > > > bakaneko (nyan.delete@this.hyan.wan) on September 29, 2013 8:59 am wrote:
> > > > > > mas (a.delete@this.b.com) on September 29, 2013 6:45 am wrote:
> > > > > > > Doug S (foo.delete@this.bar.bar) on September 28, 2013 5:09 pm wrote:
> > > > > > > > Dan Fay (daniel.fay.delete@this.gmail.com) on September 28, 2013 3:38 pm wrote:
> > > > > > > > > Michael S (already5chosen.delete@this.yahoo.com) on September 28, 2013 1:11 pm wrote:
> > > > > > > > > > Doug S (foo.delete@this.bar.bar) on September 28, 2013 11:39 am wrote:
> > > > > > > > > > >
> > > > > > > > > > > That suggests to me that it very likely costs Intel more to make Bay Trail than it costs Samsung to make
> > > > > > > > > > > A7, and when you add that to their higher gross margin the purchase price to the customer would probably
> > > > > > > > > > > be at least 50% higher
> > > > > > > > > >
> > > > > > > > > > You mean, Bay Trail Android tablets will be 50% more expensive to customers than A7-based iPad?
> > > > > > > > >
> > > > > > > > > I read it as the "customer" being the tablet maker, not the end user, but that's just
> > > > > > > > > me. In the iPhone 5S, the A7 is about 10% of the phone's manufacturing cost, and is cheaper
> > > > > > > > > than the modem and screen, and about the same price as the phone's chassis.
> > > > > > > >
> > > > > > > >
> > > > > > > > Yes, that's how I intended it. 50% more meaning $10 more than the $20 cost of A7. Not a huge
> > > > > > > > factor in terms of the final cost of the whole tablet, but not insignificant, either. Intel
> > > > > > > > could certainly make up for that with under the table "market development" money like they've
> > > > > > > > handed out for years. It doesn't appear they've played that card in the mobile market yet (or
> > > > > > > > if they have its been an utter failure) but at some point they're going to have to resort to
> > > > > > > > it to fool their shareholders into thinking they're meeting with some success versus ARM.
> > > > > > > >
> > > > > > > >
> > > > > > >
> > > > > > > You can get about 500 Bay Trail dies from a 300mm wafer assuming 80-90% yields. The
> > > > > > > raw wafer cost is about $3K as Intel does not have to pay itself a foundry margin
> > > > > > >
> > > > > > > http://www.icknowledge.com/products/icmodel.html
> > > > > > >
> > > > > > > which means each Bay Trail is about $6, add a dollar for packaging gives you $7 which means 65% GMs at $20.
> > > > > > > This is how Digitimes can report Bay Trails will be as low as $10 which I assume will be the dual-core bin
> > > > > > > rejects allowing Intel 30% margin if they have to produce
> > > > > > > non-defect extras out of the quad-core die to meet
> > > > > > > a dual-core order. Clovertrail+ will be about $7 too with
> > > > > > > Medfield and Merrifield both about $5 die cost and
> > > > > > > the 32nm Atoms will now be sold at cost to get market share
> > > > > > > against A7, A8, A9, A12 SoCs etc and they are appearing
> > > > > > > in $100-200 android devices now. Intel is price competitive
> > > > > > > *now * against much smaller ARM cores because it
> > > > > > > does not have to pay a foundry margin which bumps up wafer cost to about $5.5K (see link).
> > > > > >
> > > > > > Intel has to pay a foundry margin to its
> > > > > > own foundries unlike what you claim. It's
> > > > > > just under a different name in the books,
> > > > > > if there is any difference at all.
> > > > > >
> > > > >
> > > > > To expand on my earlier reply, it does not work like that for Intel. FABs/process
> > > > > technology comes under capital expenditure and R&D for Intel which is already
> > > > > paid for by Core revenue and profit. Atoms can piggy bank on this.
> > > >
> > > > That's silly, it doesn't matter what chip
> > > > comes first. Except you point at something
> > > > else, which you didn't explicitly state...
> > > >
> > >
> > > I did not think I needed to state that Intel's traditional x86 Core architecture already
> > > pays for its foundry and process technology out of its gross profits still leaving excess
> > > billions for buybacks/dividends/retained profit. After all it's only been Intel's business
> > > model ever since they started x86 thirty odd years ago. Sorry for assuming too much.
> > >
> > >
> > > > Also what about all the foundries which
> > > > don't need to pay to produce chips like
> > > > Haswell? Or maybe they do? :P
> > >
> > >
> >
> > People who want to investigate this further can find any view they like at
> > http://seekingalpha.com/symbol/intc
> > My opinion, scanning these articles, is that the pessimists are a lot more realistic, and
> > that the optimists are pointing to "Intel has great mobile CPUs to sell" without addressing
> > the point that's been raised here repeatedly --- the low profit in those sales.
> >
> > To address directly mas' points, look to the graph here:
> > http://seekingalpha.com/article/1121921-what-is-wrong-with-intel
> > which shows some Intel history up to the start of 2013. (I can't find a useful
> > graph that is more recent, but I would guess the trends have not changed.)
> > The points I consider notable are
> > - revenue from PCs has been flat for a while and is now turning down
> > - "other architecture segments" basically means Atom --- flat
> > and now down. MAYBE Bay Trail will turn that around?
> > - the only bright spot is the high end --- flat but no turn down. This is exactly what anyone
> > student of Clayton Christensen would predict. They're fleeing to the high end and, while making
> > desultory attempts to compete at the low end, their hearts aren't really in it. The high end
> > is fine for a while --- and that while can last for a long time if you are IBM --- but an Intel
> > concentrating on the high end is a very different Intel from the one we're used to.
> > It will be interesting to see how that high end curve reshapes
> > itself once the 2nd gen of ARM64 CPUs are released.
> >
>
> To follow up on this post, I found this interesting page:
> http://www.semiwiki.com/forum/content/2698-intel-really-delaying-14nm-move-450mm-slipping-too-euv-who-knows.html

Be very careful about semiwiki. The guy who runs it has very minimal understanding of semiconductors (let alone actual chips) and seems to have a very strong pro-foundry (TSMC in particular) agenda. I've seen a lot of stuff posted there that is wrong or simply misses the point.

As an example, there was a post explaining why the A7 was going to be quad-core...and we know how that turned out. There was also a strange post trying to explain why Samsung's 28nm process is higher performance than Intel's 22nm process.

I don't think Intel is delaying 14nm products.

> The specifically interesting graph here is the third one, which shows the value of TSMC's output at
> 2/3rds of Intel's until early 2012, at which point it rises rapidly, to overtake Intel in 2Q2013.

That's not very informative.

1. TSMC's ROI is not determined by output value, it's based on the profit they make per wafer. Even if Nvidia's revenue per wafer is $80K (say for a wafer of Tesla's), TSMC only makes $5-8K in revenue.

2. TSMC profits/revenue is split over many process nodes. In fact, the profit is probably highest for trailing edge nodes like 350nm, which are fully depreciated. The profit is probably rather low for newer nodes. And quite a few folks are very happy with older nodes (e.g., mixed signal, analog customers). In contrast, Intel makes money almost exclusively on N and N-1 nodes (e.g., 22nm and 32nm today). So the ROI decisions look quite different, especially if it becomes questionable that 'trailing edge' customers will eventually move to 20nm, etc.

3. TSMC costs and prices are going up, largely driven by more expensive techniques such as HKMG, double patterning, FinFETs, etc. Intel is better able to control these costs through co-optimization of manufacturing and design (e.g., restrictive design rules); in contrast, TSMC must support a very chaotic and difficult to manage ecosystem.

> Point is, many of the statements in this thread are hypothecated on the assumption that
> Intel's current process lead means Intel has more money to throw at the problem, and so
> will maintain that lead forever.

Forever is a long time. I don't like making projections that far into the future : )

But today and for the forseeable future, Intel certainly has far more money to invest today than any other logic foundry/IDM. More to the point, I haven't seen any evidence that the rest of the industry is closing the gap with Intel. Look at when Intel introduces new materials, everyone else takes 3-4 years to catch up (e.g., HKMG, FinFET). And everyone is still 1.5-2 years behind Intel on process shrinks.

So the distance between Intel and the industry is just as big as ever, and the significance of that gap in power, performance, cost, etc. is still huge.

>It's not clear to me that this assumption is true.
> An alternative story is that TSMC (and Samsung, for that matter) WERE poorer than Intel, meaning they
> have historically lagged, but the river of money flowing into them (because of phones and tablets)
> means they're no longer poorer, and this extra income will allow them to keep up with >Intel.

I see no evidence to really support this. First of all, Samsung is still more DRAM/NAND than logic. If you look at their logic capex, it's perhaps 1/3rd of Intel's and they also have a bit of an issue with Apple leaving for TSMC.

Also, we would expect to see more technology leadership from Samsung/TSMC in logic, which is definitely not the case.

> And all they have to do is "keep up". If they can create products every bit as good as Intel with
> nodes lagging one generation behind Intel, that's good enough --- and a lot more >profitable.

It's not clear that they are more profitable. I suspect that Intel's costs may actually be lower than TSMC.

I do agree that in the mobile space, Intel really needs to capitalize on their manufacturing advantage much more than they have in the past.

> There's also the massive dark horse in the race, name Apple's foundry game. The rumors are that Apple is
> not merely using TSMC at 20nm, but in some fashion bankrolling the fab. This would match how Apple has contracted
> with Foxconn and Pegasus, where it essentially acted as banker, buying the tooling for advanced lines, and
> engaging in some sort of rent-to-own scheme whereby for at least some period of time those lines were for
> the exclusive benefit of Apple.

TSMC doesn't do deals like that, there are already reports that they turned down such offers. Remember that Foxconn and Pegasus are low margin, low value add companies without much R&D. TSMC is quite different.

When you think about Apple's impact, it would be smart to think about how many wafers does the iPhone, iPad, etc. represent.

>If Apple does get exclusive use of TSMC 20nm for a year or two, that obviously
> helps Apple in all sorts of ways, but also has strange and unpredictable consequences for the ARM competitors
> to Apple (most notably Samsung).

That won't ever happen, it makes no sense for TSMC. Not only has TSMC never done that, but there's really no scenario where that would make sense...short of Apple buying TSMC. That's unlikely for a number of reasons.

David