TSMC signs chip deal with Apple

[SOFTengCOMPelec]

As a linux user, I deeply care about Intel's iGPU improvements. I cannot rely on binary blobs from nvidia or AMD. I also care about the reduced platform power consumption -whether at idle or load- since my desktop stays on 24X7.

But there are not enough of mine kind for Intel to bother with a desktop Broadwell, apparently.

I second this because I have had the same issues as well. Electricity (Energy) prices seem to keep on rising and rising, so, even if Broadwell is the same basic performance as Haswell, its potential power savings (from the die shrink) and other power saving improvements, would be much appreciated.
Although I really like Linux, painful incompatibilities, between various versions of Linux and various graphics cards or integrated graphics, have been a sore issue for a long time.

 

[beginner99]

How much has the desktop market segment cared about Intel's improved integrated graphics? IVB was a peculiarity in that it provided as much CPU performance improvement as it did on a new process (my guess would be it was stuff that didn't quite make the cut for SNB). If BDW is just a straight up die shrink of HSW on the CPU side then all that it has to offer is the new process tech and better integrated graphics... neither of which matter much at all for the traditional desktop market. (Which is just a way of saying yeah, I have no expectation that we're going to see higher clock speeds... mostly because there's no reason to.)

Makes far more financial sense to continue milking the 22nm fabs on the desktop side while using the expensive 14nm capacity for the more lucrative mobile market. (Well, or the markets where the process lead converts into a tangible advantage over the competition.)

Agree. What you write makes a lot of sense to me.

 

[SOFTengCOMPelec]

Agree. What you write makes a lot of sense to me.

Doesn't make sense to me.
If Intel have spent a small fortune developing Broadwell, the more market segments they can sell it in, the better the return on its huge investment.

If it's out of sheer market business greediness, to stick with ONLY Haswells for the desktop parts, then although I can understand it, I would not be impressed with Intel's marketing strategy.

 

[ShintaiDK]

Doesn't make sense to me.
If Intel have spent a small fortune developing Broadwell, the more market segments they can sell it in, the better the return on its huge investment.

If it's out of sheer market business greediness, to stick with ONLY Haswells for the desktop parts, then although I can understand it, I would not be impressed with Intel's marketing strategy.

Or its because, just like with 32nm and 22nm that Intel can only apply the new process node in the beginning for ~2 segments.

As others already have said. Lynnfield was another example of 45nm quadcores, while mobile, dualcore and servers/workstations got 32nm.

With 14nm the focus is mobile and smartphones/tablets. Its mobile Broadwell and Airmont that get first 14nm access. Desktop and servers/workstations comes next when capacity is expanded.

 

[SOFTengCOMPelec]

Or its because, just like with 32nm and 22nm that Intel can only apply the new process node in the beginning for ~2 segments.

As others already have said. Lynnfield was another example of 45nm quadcores, while mobile, dualcore and servers/workstations got 32nm.

With 14nm the focus is mobile and smartphones/tablets. Its mobile Broadwell and Airmont that get first 14nm access. Desktop and servers/workstations comes next when capacity is expanded.

That does make sense.
Also, in relation to what you have just said, the smaller and simpler chips are probably a better proposition for the early days of a new process node, such as 14nm. (Presumably) The bigger chips need good yields, which won't necessarily be initially achieved in brand new process nodes.

I guess it's also possible, that the potentially higher clock speeds, which desktop parts need (want), take longer to get right on new process nodes. Mobile stuff tends to be lower clock rates, but benefits from the reduction in power consumption.

 

[Khato]

(Presumably) The bigger chips need good yields, which won't necessarily be initially achieved in brand new process nodes.

That's almost certainly part of it. In the past you almost always wanted to transition to the new process right away because the reduced die sizes more than offset the drop in yields and you got better performance as an extra bonus. Now the majority of die size reduction is set aside to increase the area devoted to integrated graphics and instead of better performance you get better efficiency. The result being that while before your manufacturing cost per die likely stayed roughly even when switching to the new process with a higher defect rate it's almost certainly going to cost Intel more to manufacture a BDW chip than it will HSW... and if the desktop version offers no marketable advantage over HSW then why should they offer it?

Regardless, we're all just guessing at the exact reason. But it's pretty much a given that there's a good business reason for it.

 

[SOFTengCOMPelec]

That's almost certainly part of it. In the past you almost always wanted to transition to the new process right away because the reduced die sizes more than offset the drop in yields and you got better performance as an extra bonus. Now the majority of die size reduction is set aside to increase the area devoted to integrated graphics and instead of better performance you get better efficiency. The result being that while before your manufacturing cost per die likely stayed roughly even when switching to the new process with a higher defect rate it's almost certainly going to cost Intel more to manufacture a BDW chip than it will HSW... and if the desktop version offers no marketable advantage over HSW then why should they offer it?

Regardless, we're all just guessing at the exact reason. But it's pretty much a given that there's a good business reason for it.

A bit like an extension of the TICK/TOCK approach, then.

The idea of the TICK/TOCK (as I'm sure most know anyway) is to separate out die shrink changes, and fundamental cpu core architecture re-design.

So, maybe it's now more of a:

• TICK/TOCK-DIE-SHRINK-FOR-MOBILE
• TICK/TOCK-DIE-SHRINK-FOR-DESKTOP(this version might not happen)
• TICK/TOCK-DIE-SHRINK-NEW-ARCHITECTURE-FOR-MOBILE&DESKTOP
• TICK/TOCK-DIE-SHRINK-NEW-ARCHITECTURE-FOR-XEON&E-TYPE(SOCKET2011++)

With each of the above steps being something like 6..18 months apart.

 

[ShintaiDK]

That does make sense.
Also, in relation to what you have just said, the smaller and simpler chips are probably a better proposition for the early days of a new process node, such as 14nm. (Presumably) The bigger chips need good yields, which won't necessarily be initially achieved in brand new process nodes.

I guess it's also possible, that the potentially higher clock speeds, which desktop parts need (want), take longer to get right on new process nodes. Mobile stuff tends to be lower clock rates, but benefits from the reduction in power consumption.

Broadwell mobile and desktop is the same size. You could even say the mobile is bigger like with most Haswell and IB.

 

[SOFTengCOMPelec]

Broadwell mobile and desktop is the same size. You could even say the mobile is bigger like with most Haswell and IB.

I was more thinking of stuff like Silvermont. But, yes, I agree with you.

Broadwell mobile chips, would make great mini-ITX motherboards (BGA etc), as the mini-ITX format would like the low TDP broadwell versions, as it would help make it a very quiet or silent, yet powerful and very energy efficient computer.

Unfortunately, I fear that Intel would class the above as "desktop".

 

[AtenRa]

That's almost certainly part of it. In the past you almost always wanted to transition to the new process right away because the reduced die sizes more than offset the drop in yields and you got better performance as an extra bonus. Now the majority of die size reduction is set aside to increase the area devoted to integrated graphics and instead of better performance you get better efficiency. The result being that while before your manufacturing cost per die likely stayed roughly even when switching to the new process with a higher defect rate it's almost certainly going to cost Intel more to manufacture a BDW chip than it will HSW... and if the desktop version offers no marketable advantage over HSW then why should they offer it?

Regardless, we're all just guessing at the exact reason. But it's pretty much a given that there's a good business reason for it.

Traditionally you start with the big die chips because they are the faster chips that comes at a higher price giving you the higher margins. You offset your lower yields with premium prices (see AMD HD7970 and GTX680). Then when you have higher yields and you have made some of the manufacturing cost through the high priced parts (HD7970) then you can start producing smaller high volume lower margins chips. We see this every year with Intel(CPU) and AMD/NVIDIA(GPU).

Now, the fact that Intel will release a smaller die, lower frequency, low power chip first (Mobile U and Y) with a new process might be due to the process it self. Intel may very well have started this round(14nm) with the low power process first, with the high performance 14nm process following later(Desktop 2015).
Knowing that Intel is after the mobile and especially the low power Smart Phones market with Mobile ATOM, it looks more and more that they will use the Low Power 14nm process first.

Could be wrong though

 

[scannall]

Wow, the comments have sure wandered far afield, being the OP was talking about Apple and TSMC.

It will be interesting to see what this does to Samsung's foundry business. I'm betting that TSMC is in a great place to gain the lions share of that business, since they have the all the big 'anchor' clients like Apple, Qualcomm, nVidia, AMD, Broadcom etc. Intel doesn't count, since they do very little outside work. GloFlo isn't really a competitor at this point in time.

 

[Kenmitch]

Samsung - Obviously this takes billions off the table. Even though Samsung's profit margin may not be much producing Apple's "A" series of ARM processors, that's still a huge blow. These things tend to have a trickle down effect that could affect more than just the dollars Samsung was getting from Apple. Things like higher procurement cost due to lower purchasing volumes can't be ignored.

Samsung could always just start cranking out more wonder ram I'd think. Most likely with todays market prices of DDR3 it would be way more profitable anyways.

 

[Sheep221]

TSMC is not only one semiconductor foundry beside Intel, infineon, STmicroelectornics, globalfoundries, texas instrument and more have fabs too.

 

[Kenmitch]

TSMC is not only one semiconductor foundry beside Intel, infineon, STmicroelectornics, globalfoundries, texas instrument and more have fabs too.

The following was also stated in the article

The sources did not identify whether TSMC will be the sole supplier of these Apple-designed chips

 

[sm625]

65nm Q1 2006.

45nm Q1 2008.

32nm Q1 2010

22nm (one quater late) Q2 2012

14nm on time would be Q1 2014. If the delay was a one time thing then Q2 2014, if the delay is additive then its Q3 2014.

 

[mrmt]

65nm Q1 2006.

45nm Q1 2008.

32nm Q1 2010

22nm (one quater late) Q2 2012

14nm on time would be Q1 2014. If the delay was a one time thing then Q2 2014, if the delay is additive then its Q3 2014.

One quarter on the span of 8 years and four (!) nodes. Not bad, is it?

 

[Exophase]

65nm Q1 2006.

45nm Q1 2008.

32nm Q1 2010

22nm (one quater late) Q2 2012

14nm on time would be Q1 2014. If the delay was a one time thing then Q2 2014, if the delay is additive then its Q3 2014.

Intel released 45nm November 2007, 32nm January 2010, 22nm April 2012. It's about 26 then 27 months.

 

[bryanW1995]

From 3.5B to 4.5B, yes thats fantastic until you look on the competition. TSMC grew from 14.6B to 17B.

TSMC is on 20nm, when GloFo is on 28nm. A node TSMC already milked leaving discount products to GloFo.

I haven't been paying much attention to this lately, but, last time I checked, 31% > 16.4%. And, at least a few years when I was actually paying attention to the fab wars, the consensus was that TSMC was really struggling with delivering 28nm gpus in a timely fashion.

I hope Apple knows what they are doing. They seem to be making more and more mistakes lately now that Jobs is gone.

 

[Exophase]

I haven't been paying much attention to this lately, but, last time I checked, 31% > 16.4%. And, at least a few years when I was actually paying attention to the fab wars, the consensus was that TSMC was really struggling with delivering 28nm gpus in a timely fashion.

I hope Apple knows what they are doing. They seem to be making more and more mistakes lately now that Jobs is gone.

No matter how much TSMC may have struggled with 28nm devices early on they still brought them to market over a year earlier than GF or Samsung. They even came out before products on Samsung's 32nm - that would be what current Apple SoCs are made on.

If you think about it, it's pretty natural for Apple to be moving away from Samsung's fabs, even if you were to ignore the legal and competitive friction between the two companies. I'm not aware of a single large volume product (or anything at all, really) made by Samsung's fab that isn't Samsung's or Apple's. So it's hardly a popular default choice over TSMC. The only reason Apple is using them now is because they started out using standard Samsung SoCs, then SoCs customized by Samsung, then gradually transitioned to their own designs. Along the way they acquired Intrinsity who was previously closely affiliated with Samsung, therefore they'd have had more experience sticking with Samsung. But in the long term it made more sense to migrate somewhere else.

 

[RampantAndroid]

I haven't been paying much attention to this lately, but, last time I checked, 31% > 16.4%. And, at least a few years when I was actually paying attention to the fab wars, the consensus was that TSMC was really struggling with delivering 28nm gpus in a timely fashion.

I hope Apple knows what they are doing. They seem to be making more and more mistakes lately now that Jobs is gone.

Yes, and if I only have 1 and get 2, that's 100%! And that's WAY more impressive than Joe over here going from having 1,000 to having 1,500. Because that's only 50%.

Percentages are JUST THAT. Numbers that alone mean nothing.

Exophase: I thought all of Qualcomm's stuff was fabbed by Samsung?

 

[Exophase]

Exophase: I thought all of Qualcomm's stuff was fabbed by Samsung?

Qualcomm's SoCs are fabbed entirely at TSMC. I bet they're TSMC's largest customer.

 

[Idontcare]

Qualcomm's SoCs are fabbed entirely at TSMC. I bet they're TSMC's largest customer.

Any fabless company that is serious about competing against Samsung products in the retail space know better than to hand over their roadmap plans for ICs to be produced in 2-3yrs time to the foundry side of Samsung. Some make a mistake or two in their corporate learning curve, but they learn quickly.

 

[ShintaiDK]

Samsung, without Apple, is down to a single billion$ MPU revenue.

 

[Idontcare]

From 3.5B to 4.5B, yes thats fantastic until you look on the competition. TSMC grew from 14.6B to 17B.

TSMC is on 20nm, when GloFo is on 28nm. A node TSMC already milked leaving discount products to GloFo.

I haven't been paying much attention to this lately, but, last time I checked, 31% > 16.4%. And, at least a few years when I was actually paying attention to the fab wars, the consensus was that TSMC was really struggling with delivering 28nm gpus in a timely fashion.

I hope Apple knows what they are doing. They seem to be making more and more mistakes lately now that Jobs is gone.

Bryan, don't get lost in the percentages, that was a red-herring. The relevance of comparing the revenue numbers (and growth in them) comes from the issue of how to pay for things going forward.

That extra $1B that GloFo has in revenue is not going to pay for a new fab, nor is it going to pay for the development of a feature-competitive 10nm node on a competitive timeline either.

That extra $2.4B that TSMC made in revenue though will pay for a substantial portion of capex in a new fab, as well as deliver a more feature-complete (clockspeeds, power, density, etc) 10nm node on a timeline that will rock its competitor's customers.

Breaking things down to percentages is fine if you just want to do math that is technically correct but meaningless all the same. But if you are serious about asking yourself big questions like "who is getting to 10nm first, who will have 10nm volume availability first, who will have a better quality 10nm node?" etc then you have to look at the absolute numbers because that is what pays those bills.

 

[Exophase]

Any fabless company that is serious about competing against Samsung products in the retail space know better than to hand over their roadmap plans for ICs to be produced in 2-3yrs time to the foundry side of Samsung. Some make a mistake or two in their corporate learning curve, but they learn quickly.

People like to say this but I don't see any evidence that Samsung ever stole anything submitted to their foundries. Samsung Mobile (a fairly separate entity) got blasted for making devices with physical design too similar to iPhones. That has nothing at all to do with the SoCs. I don't think Apple had to hand over anything of their higher level product roadmaps to Samsung's foundries to get their chips made.