Sony to leave three-way microchip development pact with NEC and Toshiba

[3dilettante]

You and Capeta are viewing things in way too PC-centric a manner, as if the rest of the world of semiconductors is 'meh' because we don't read about it on tech sites every day. You haven't answered my question yet Deusp as to why companies like AMD and IBM need advanced fabs, so I'm still waiting on that - personally I don't think there's any logic that can be applied there that will be exclusive to those companies. Why shouldn't AMD just outsource everything to Chartered and stop making their own fabs? Indeed, it will be interesting to see what node the chips inside the iPhone do indeed use - Capeta guesses 90nm... perhaps. But it's a lot more likely to be 65nm than 130nm for instance, agreed?

I can't speak to the entire range of reasons why a company would decide one way or the other, but IBM and AMD do have a more pressing demand for a very specific subset of high-performance processes.

I've seen discussions and articles indicating that high-end MPU processes are very targeted, which makes them less suitable for analog and specialized products.
Analog doesn't really scale anyway, but it's a given that any non-MPU product on a high-end process like AMD's would be leakier, hotter, and less reliable (assuming it can work as designed). Analog and mixed signal hardware is casually brushed aside on such processes.

IBM and AMD are also subject to a constraint that circuit performance at the next process node improve over the last. TI's dropping out of the scaling race in processes like those used by Sun's processors is a symptom of the end of this trend.
One need only look at Intel's 45nm process to see that heroic efforts are needed to maintain a performance improvement.
For a current example, look at the currently nonexisted clock speed gains AMD got from shifting to 65nm production. The highest clocking chips are 90nm.
In other designs, the already subnominal density scaling high-speed logic gets from process transitions is getting worse, because the densest circuits do not perform as well.

65nm may be what will be known in the future as the end of guaranteed performance scaling by process transition.

CCDs don't need to worry about maintaining gigahertz clock speeds.
Non MPUs can benefit from less performant processes, because they need only care about density or power consumption.
(edit to word this better: their power and density constraints restrict their need for circuit performance to well within the range offered by a less performance-driven process)

While future nodes will be expensive, future high-performance nodes are even more so.
Worse, high-performance nodes cater to a very limited market compared to the rest of the IC world.

This is also a reason why IBM's cooperative strategy may not win against Intel in the MPU space. Every partner has a certain ideal process, which leads to a more general and less optimized solution.
It's one of the reasons why foundry chips clock so far below in-house fabs. The more general needs of a TSMC foundry's client range means a number of circuit types and safety margins that can be removed or smoothed over with an Intel process must be present, even if at the cost of longer timings.
Realworldtech had an article that put timings for a foundry at 2-3x those of an Intel one.

 

[Carl B]

3dilet, thanks for the insights. Within the framework you set up...

CCDs don't need to worry about maintaining gigahertz clock speeds.
Non MPUs can benefit from less performant processes, because they need only care about density or power consumption.
(edit to word this better: their power and density constraints restrict their need for circuit performance to well within the range offered by a less performance-driven process)

Well in terms of the thread as far as it relates to Sony/NEC/Toshiba and the CMOS process shrinks, I think this would be exactly the angle - simply density and power characteristics vs performance-related. The equation basically boiling down to more chips/wafer, higher margins/chip, and better power and physical profiles for the CE devices that use them.

This is also a reason why IBM's cooperative strategy may not win against Intel in the MPU space. Every partner has a certain ideal process, which leads to a more general and less optimized solution.

It's one of the reasons why foundry chips clock so far below in-house fabs. The more general needs of a TSMC foundry's client range means a number of circuit types and safety margins that can be removed or smoothed over with an Intel process must be present, even if at the cost of longer timings.
Realworldtech had an article that put timings for a foundry at 2-3x those of an Intel one.

This is an interesting point to raise... I think we all know Intel is simply above the rest in terms of what they can achieve on the manufacturing side, but still even with the disagreements in this thread as to the industry in general, I think all of us view IBMs fab coalition as a sort of rebel alliance against a galactic empire, and would like to see it do well. Yet you raise the point that since IBMs internal demands are so low, that the work they do in process development for their various and disparate coalition partners actually may work to dilute the process technology in its ability to compete.

Interesting point also on the foundries and where their processes' stand in relation to internal efforts geared towards high-performance; didn't realize the performance spread was that great. Nakagawa's statement that 45nm Cell production may go to foundries would obviously require a pretty serious process shift then to run a 3.2GHz, not to mention the move from SOI to non. It'll be interesting to see how the move to 32nm effects the next gen console chips as well, and what paths they all decide to take.

 

[3dilettante]

Well in terms of the thread as far as it relates to Sony/NEC/Toshiba and the CMOS process shrinks, I think this would be exactly the angle - simply density and power characteristics vs performance-related. The equation basically boiling down to more chips/wafer, higher margins/chip, and better power and physical profiles for the CE devices that use them.

Sony's opting out of this particular agreement may indicate a shift in priorities that separates the partners. Not knowing what the targets are for each individual company, it might be that Sony no longer has the same process targets as the other two.

If the process they are working towards is not something that fits well with all the partners, they have less individual incentive to remain.

There may be some kind of curve to the cost/benefits analysis, depending on the product mix each company wants.

If Sony still wants to make Cell, it's not likely a process devoted purely to density or power will be satisfactory.
Future foundry performance processes may not beat Intel's timings, but they would be better than plain bulk processes.

Perhaps Sony is still looking for decent circuit performance, but doesn't think it will have the volumes to justify an in-house high performance line. A foundry could fill the volumes needed to justify such a line with multiple customers.
It wouldn't beat Intel, but unless Sony is thinking about entering the CPU market, it doesn't need to.

It's IBM that needs to worry about the part of its business that needs the top-end performance.

This is an interesting point to raise... I think we all know Intel is simply above the rest in terms of what they can achieve on the manufacturing side, but still even with the disagreements in this thread as to the industry in general, I think all of us view IBMs fab coalition as a sort of rebel alliance against a galactic empire, and would like to see it do well. Yet you raise the point that since IBMs internal demands are so low, that the work they do in process development for their various and disparate coalition partners actually may work to dilute the process technology in its ability to compete.

The downsides to this approach may be seen in IBM and AMD's slow 65nm ramps.

Both AMD and IBM have an incremental process transition policy. With AMD especially, we can see how they've stated that the initial 65nm transistors performed the same as the 90nm ones. Refreshes over time were supposed to push the 65nm performance over the optimized 90nm process.

Perhaps this approach works better when there are many separate initiatives that have to be integrated in a way that fits each company's goals.

Intel's wholy in-house processes, on the other hand, have far fewer process refreshes. During the Athlon era, this lead to gaps of time where AMD's circuit design was more competitive with Intel's.
AMD's fallen off track quite a bit since then.

Interesting point also on the foundries and where their processes' stand in relation to internal efforts geared towards high-performance; didn't realize the performance spread was that great. Nakagawa's statement that 45nm Cell production may go to foundries would obviously require a pretty serious process shift then to run a 3.2GHz, not to mention the move from SOI to non. It'll be interesting to see how the move to 32nm effects the next gen console chips as well, and what paths they all decide to take.

There's Chartered. It already has an SOI process available, as Microsoft can attest.
By the time 45nm Cell comes around, there may be more options.

 

[standing ovation]

I see them as naturally complementary though - imagine it like this, you have the car industry with its own in-house design and manufacturing capabilities; do you believe maintaining one comes at the expense of the other?

Yes, simply because there isn't enough money to go around. So you end up saying, "What's our priority: An idea that works in theory or reality?"

 

[Carl B]

If Sony still wants to make Cell, it's not likely a process devoted purely to density or power will be satisfactory.
Future foundry performance processes may not beat Intel's timings, but they would be better than plain bulk processes.

Well I wasn't speakign to Cell - obviously Cell is a different animal altogether. But TV and digital sensor ICs was more where this particular process would have been targeted. Sony does a brisk CMOS sensor business and all projections of theirs have always shown demand as exceeding their capacity to supply.

There's Chartered. It already has an SOI process available, as Microsoft can attest.
By the time 45nm Cell comes around, there may be more options.

Chartered is the only one of the three though, and it's pretty small. On the plus side their SOI process is basically IBMs, which is basically Sony's, so the transition would be eased... but then again we're talking 45nm and I'm not sure Chartered 45nm SOI plans are well known at this point. I don't have anything against them of course either, but watching AMDs on-again/off-again fab arrangements with them over the years just sort of gets a flag in my mind, everything else otherwise being shrouded in mystery.

 

[Carl B]

Yes, simply because there isn't enough money to go around. So you end up saying, "What's our priority: An idea that works in theory or reality?"

Well fair enough, but we don't knwo what reality is or is not at this point. Sony's withdrawn from this fab alliance, but at times I think this thread is in the context of "Sony announces fab outsourcing to foundries," which, hasn't happened yet. Let's wait for that to happen and then you can tell me reality set in.

As of yet, they are still not commited to one path or another. Afterall, these costs have been known for some time, yet exactly one year ago they announced an alliance to pursue an internal effort... so things are still in flux over there to an extent, with a push/pull we're not privy to.

 

[3dilettante]

Well I wasn't speakign to Cell - obviously Cell is a different animal altogether. But TV and digital sensor ICs was more where this particular process would have been targeted. Sony does a brisk CMOS sensor business and all projections of theirs have always shown demand as exceeding their capacity to supply.

I see, I hadn't mentally transitioned from the discussion on high-end processes.

Are there physical concerns about sensor reliability at the finer geometries?

Sony could always increase fab capacity at the current process and defer on a process transition. A timetable change alone could lead to the dissolution of a partnership that had a different deadline.

Chartered is the only one of the three though, and it's pretty small. On the plus side their SOI process is basically IBMs, which is basically Sony's, so the transition would be eased... but then again we're talking 45nm and I'm not sure Chartered 45nm SOI plans are well known at this point. I don't have anything against them of course either, but watching AMDs on-again/off-again fab arrangements with them over the years just sort of gets a flag in my mind, everything else otherwise being shrouded in mystery.

It beats me. AMD's on-again/off-again product leadership might be to blame.

With Intel trouncing them so thoroughly, the lower-end chips Charter would most likely produce would have dropped in price to the point that Charter's extra cost would have made them infeasable. Additional engineering for Charter's process to get acceptable turnout for a complex x86 would have made the break-even point even harder to hit.

Since AMD's 65nm process seems incapable of yielding many high-end chips months into its transition, its low-end volume production for the sake of fab utilization would smack right into what Charter would have been offering.

 

[Fox5]

We've ignored Samsung here since they make memory chips and not ASIC. Unlike ASICs, designing memory chip cost very little and hence process is nearly everything. They also need specialized fabs too. Samsung can justify it's internal fabs but most CE companies can't.

Uh, memory companies make up a large amount of that list of top fabs, and a lot of them produce other types of chips as well.

Anyhow, any company who produces enough chips that the profit per chip of their own fabs not only exceeds what is available by outsourcing but recoups the cost of the fabs could benefit from having their own fab.

I think I get what is being referred to by saying only AMD, Intel, and IBM need their own fabs, that only they generate the kind of revenues from a single product needed to outperform the industry in fabs but....that's really only true for Intel. AMD and IBM aren't far ahead of TSMC and Chartered and the rest, and as I recall, both were stuck on 90nm for quite some time because of cost reasons, not because they couldn't do 65nm.

From your logic, Intel is the only company in the world that needs its own fabs (and apparently memory companies).
I guess everyone who needs advanced fabs could just pony up to IBM, but that could get awfully expensive if too many companies decide they need the extra half mile IBM offers over TSMC.

Intel's and IBM's, with everyone's fabs based off of one of these two companies' technology. If this is endgame result, then whether you go foundry or internal the end result may be exactly the same.

My prediction:
1st place: Intel
2nd place: IBM
3rd place: Combined cooperation between Taiwanese and possibly Chinese efforts...heck, let's just make it Asia in general, anyone who doesn't team up with IBM is probably going to team up with each other

And there will still be additional fab efforts beyond that.
Oh wait, that's pretty much how it looks today.
Longer term, look for massive investments by China with current 2nd and 3rd tier players.