Intel to Outsource 14nm Chip Production to TSMC Due to Tight Supply

[Nelly]

Intel is encountering tight 14nm process production capacity in-house, and is looking to outsource part of its 14nm chip production to Taiwan Semiconductor Manufacturing Company (TSMC), according to industry sources.

Intel intends to give priority to its high-margin products mainly server-use processors and chipsets amid its tight 14nm process capacity, and therefore plans to outsource the production of its entry-level H310 and several other 300 series desktop processors to TSMC, the sources indicated.

Intel has seen its overall 14nm chip supply fall short of demand by as much as 50%, the sources said. Outsourcing has become the only and appropriate choice for Intel since the company is unlikely to build additional 14nm process capacity, the sources noted.

TSMC is already a contract manufacturer of Intel for SoFIA-series handset SoC chips and FPGA products, and makes Intel's baseband chips for use in the iPhone, the sources said.

Motherboard makers expect the tight supply of Intel's 14nm chipsets to ease by the end of 2018.

Market observers believe that Intel's tight capacity for 14nm stems from its delay in advancing to 10nm. Intel originally planned to enter mass production of its 10nm Cannon Lake processors in 2016, but has been pushing back the schedule.

Intel's latest update is that its 10nm chips will not be ready for commercial production until the fourth quarter of 2019.

Intel had been adopting the so-called "Tick-Tock" model under which every microarchitecture change is followed with a die shrink of the process technology, but the "Tick-Tock" production cycle started slowing down in 2014. Intel's 14nm manufacturing node has become the "longest-lived" process technology in the chip giant's history, the observers noted.

Intel has declined to comment on the reported move to outsource 14nm chips to TSMC.

Source

 

[Armageus]

I hear global foundries may have some 14nm capacity

 

[SiDeards73]

I hear global foundries may have some 14nm capacity

Lol they said they was going to focus on improving 14/12nm for smaller products, so chipsets etc would probably be ideal

 

[4K8KW10]

If in 2016 someone said that two years later the situation for Intel would be so dire, many would call him a lunatic.
Today, the worst nightmares for Intel are all reality and happening.

 

[Rroff]

I wonder how long before they are going to TSMC for sub 14nm CPU production

 

[AndreiD]

@4K8KW10 Check Intel's quarterly reports if you think they're in a 'dire' situation, they're selling everything faster than they can produce it.

And I highly doubt they're going to use TSMC for any processors, the gigantic 22nm fab they have in Israel is getting upgraded to 10nm and that has a 150,000 wafer output per month. As a reference, AMD uses the GloFo's New York foundry for Ryzen production, and all of its Ryzen line is ~60K wafers per month.

Though it's still an image hit for Intel to have to resort to TSMC for production, even though it's chipsets. Shows how badly their 10nm transition has been planned.

 

[drunkenmaster]

@4K8KW10 Check Intel's quarterly reports if you think they're in a 'dire' situation, they're selling everything faster than they can produce it.

And I highly doubt they're going to use TSMC for any processors, the gigantic 22nm fab they have in Israel is getting upgraded to 10nm and that has a 150,000 wafer output per month. As a reference, AMD uses the GloFo's New York foundry for Ryzen production, and all of its Ryzen line is ~60K wafers per month.

Though it's still an image hit for Intel to have to resort to TSMC for production, even though it's chipsets. Shows how badly their 10nm transition has been planned.

150k 22nm wafers doesn't translate to 150k 10nm wafers. Going from primarily single patterning 22nm to a quad patterning 10nm changes the time wafers take to make and increases the number of production stages massively. Even 14nm with dual patterning has significantly dropped capacity. I'm not sure where you got the 150k from at all tbh as I see Intel make no claims of those numbers. Regardless, Intel's fab 28 has 200k square feet of clean room space, Global's Fab 8 has 210k square feet of clean room space.

Sometimes as 200mm wafers were standard in the industry for a long time and is still used in large volume by TSMC and many others the 300mm numbers get standardised against 200mm fabs. Global was rated at 60k 300mm wafers a month on construction and this has since been expanded to closer to 80k a month. LIkewise they rated this at the equivalent of 135k 200mm wafers a month.

The likelyhood is Fab 8 is close or a bit bigger than fab 28. That doesn't change that Intel has multiple fabs and huge capacity, but they also had so much excess capacity on 14nm that they shut the doors on a brand new fab they built waiting for future demand with the intention of shutting down older less capable facilities at the same time.

https://newsroom.intel.com/editorials/fab-42-recent-announcement/

It's quite hilarious how the timing has worked out. They built fab 42 but found that less demand meant they didn't want to spend billions on 14nm equipment just to move to 10nm or beyond. AS it happened they planned to wait all the way to 7nm to use it. So they are struggling badly for 14nm capacity, have a completely empty fab without even any equipment in it (maybe some test 7nm/euv stuff but certainly not full up). Somehow they've switched too much existing 14nm capacity to 10nm before 10nm was ready, have loads of empty fab space and are making too few chips to meet demand all at the same time. Concurrently they were waiting for 7nm for fab 42 while they can't make 10nm work at all. There's a reason the CEO got fired, the handling of this fiasco is laughable. Icelake/10nm delayed by ~4 years IF it comes out late 2019, spending billions on fabs they shut the doors on waiting for future equipment. Screw up a 2 year industry lead on process node which was the single biggest advantage they had over everyone else and sit on the same architecture for years which lead to the problem of process being their only big advantage any more.

As for dire situation, for a company like Intel losing sales of billions because you can't make chips due to your own screw ups is dire... hence a CEO getting canned, hence hiring a bunch of new execs to head up every different major chip design department because your current architecture hasn't moved forward, hence a huge stock drop in recent months and that is all before the upcoming year they'll be a full node behind AMD who will have a new architecture which takes another step forward.

To you and me, making 10 billion one year and 8 billion the next is like... dude, I made 8 billion.... wooooooo. To a publicly traded company making less money than you should be making is borderline crisis level situation. But really the issue isn't how it is now, it's how they are expecting demand to drop next year and profitability to crash next year. They'll still be profitable but as above, a reduction in profit is bad to such companies.

When Apple had reduced growth in profit everyone panicked, that's right, a reduction in growth, profit was still higher than the previous year but the growth in profit slowed and wallstreet had a panic attack.

More than that, what happens next year will change the fight for the next 5 years. If AMD gained 1% server market share and stuck there for 5 years, Intel don't care. The massive disparity there will be between Intel and AMD in server could mean that AMD have a 15-20% market share by the end of 2021 and once they have that market share and without a process advantage to beat AMD with there is no telling that they'll ever win that back. Basically their screw ups are opening the door for AMD and once they get their foot in the door it may be impossible to shut it again.

 

[AndreiD]

I remember reading a semiengineering or wikichip article regarding fab 28 but I can't seem to find it, so I guess the 150K wafer number should be taken with lots of grains of salt.
But even with SAQP I think the likelihood that Fab 28 will output more wafers than GF's Fab 8 is pretty high, that fab has went through several rounds of investment to expand capacity/upgrade production. I can't find recent numbers for what clean room space Fab 28 is at right now sadly, so you're right that it's difficult to compare the two, only more palpable numbers we have to go on are the investment ones, and in that regard Fab 28 with its most recent planned expansion does dwarf GF's Fab 8.

I think their continued 10nm delays have forced them to keep on using 14nm for most of their product stack which got them in this situation. But to be fair this isn't the first time they outsourced production to TSMC, a few networking and FPGA SKUs are also TSMC manufactured.

I'm not sure if Fab 42 is empty or without equipment, I thought it was already producing 14nm. But I agree, Intel's management handled this pretty poorly.

But they aren't making less money than they should be making, they've constantly been beating expectations and they've already raised guidance for Q3 because demand is so high. You'd expect all of this increased competition from both AMD and ARM chips in data center will impact their sales, but it hasn't really. And I also don't see the reduction in growth you're talking about, maybe apart from their client computing group? But their other divisions, especially data center, have been beating expectations.

You're right that all of the 10nm screw ups are opening the door for AMD and in my opinion that's a good thing since more competition is better for everyone and leads to more innovation. I also hope AMD does get to 15~20% of data center market share by end of 2021, but I honestly don't see it, at least not until more Epyc 2 details crop up.

 

[CAT-THE-FIFTH]

Another factor is also that their newer products are having larger and larger dies too,so I would expect less chips per wafer and if you add chipsets being made on 14nm now you can see the problem.

 

[Rroff]

Though it's still an image hit for Intel to have to resort to TSMC for production, even though it's chipsets. Shows how badly their 10nm transition has been planned.

To be fair for incidental stuff i.e. certain peripheral controllers they've used TSMC before and were in talks to use them again before this.

As for dire situation, for a company like Intel losing sales of billions because you can't make chips due to your own screw ups is dire... hence a CEO getting canned, hence hiring a bunch of new execs to head up every different major chip design department because your current architecture hasn't moved forward, hence a huge stock drop in recent months and that is all before the upcoming year they'll be a full node behind AMD who will have a new architecture which takes another step forward.

Intel can afford it though - not indefinitely - but certain tank it for awhile so it isn't dire dire - though they've canned atleast 3 people one way or another over it (probably) so far so it certainly isn't insignificant.

 

[AndreiD]

Apparently fud from Digitimes, it's been denied by Intel: https://www.pcgamesn.com/intel-denies-tsmc-outsourcing

 

[SiDeards73]

Apparently fud from Digitimes, it's been denied by Intel: https://www.pcgamesn.com/intel-denies-tsmc-outsourcing

Yet prices on intel chips are increasing? lol... Of course Intel dont want everyone to know they have capacity issues, that would devalue their share prices further...

 

[AndreiD]

They didn't deny that they can't meet demand, but they did deny that they're outsourcing to TSMC by saying they're investing in more 14nm production lines in their own fabs.
Increased pricing is due to demand being so high, they're in a position to do that, albeit not great for consumers.

 

[EsaT]

But even with SAQP I think the likelihood that Fab 28 will output more wafers than GF's Fab 8 is pretty high...

But they aren't making less money than they should be making, they've constantly been beating expectations and they've already raised guidance for Q3 because demand is so high. You'd expect all of this increased competition from both AMD and ARM chips in data center will impact their sales, but it hasn't really. And I also don't see the reduction in growth you're talking about, maybe apart from their client computing group? But their other divisions, especially data center, have been beating expectations.

With AMD moving high end production to TSMC, GF's capacity won't matter that much.
TSMC simply dwarfs other "open for everyone to use" fabs:
http://www.icinsights.com/news/bulletins/tsmc-continues-to-dominate-the-worldwide-foundry-market/
And I wonder how much of Intel's factory space is now useless with 10nm being broken/in development hell.
Kinda suspect it's basically FUBAR and won't come out with all the "goodies" originally advertised.

And it's exactly that datacenter/server business which would be taking hit, if market didn't have existing Intel bias.
Meltdown/Spectre have their biggest performance penalty there.
And those markets have little use for highest single thread performance, with emphasis being on number of cores/threads.
Which is were AMD's modular design shines over Intel's expensive to manufacture huge monolithic (or should we say megalithic) chips.

Zen2 and 7nm is just going to increase that difference even further.
With 8 core chip becoming lot smaller than currently, or then core count increasing.
While Intel can neither increase core counts from current without cost skyrocketing, nor lower pricing without cutting into profits.
Though with their bank account size Intel can afford low profits for some years, even if investors won't like that.

If in 2016 someone said that two years later the situation for Intel would be so dire, many would call him a lunatic.
Today, the worst nightmares for Intel are all reality and happening.

Wouldn't have myself believed problems Intel is now facing.
Question is simply how far brand blindness will carry Intel, before these problems fully show their effect.

 

[LePhuronn]

Question is simply how far brand blindness will carry Intel, before these problems fully show their effect.

In the tiny corner that is our consumer market, I don't think it will ever make a difference. A lot of the vulnerabilities don't affect us enough to measure performance differentials (unless you're an enthusiast), and it would take an Athlon-style beatdown for the myopic, clueless and loyalists to ever consider AMD to be better. And even then mindshare and loyalty will win out (coming from the guy who went Socket 478 Prescott out of ignorance).

HEDT and workstations? I'd say Threadripper is starting to sway the market, but until (for example) all 32 cores on the 2990WX are fully maxed in software, Intel will probably still get the nod because the price difference can be recouped by reduced workload times that come from Intel's faster cores.

Datacenters and all the top-end stuff is where Intel can't get away with loyality/blindness any more. Meltdown and Spectre mitigations have already given tangible performance metrics and even if Intel could make hardware changes soon (Cascade Lake I think?), they're jammed on 14nm++++++++++++ so you're still getting chunky, hungry and hot processors for the foreseeable future. Compared with the different approach EPYC has taken, and finally getting proper certifications, all this is empirical evidence for those who need to research and tune every aspect of a platform's pros and cons. And the ball is rolling big time: Dell, HP, DropBox, Microsoft, Yahoo Japan, Packet, Baidu, Cisco, Cray, all have jumped on EPYC.

tl;dr: effects won't be felt in the consumer space, effects will foster decision making in the HEDT/workstation space, effects are already felt in the server space and it's shafting Intel.

 

[EsaT]

In the tiny corner that is our consumer market, I don't think it will ever make a difference. A lot of the vulnerabilities don't affect us enough to measure performance differentials

Indeed that max few percents difference doesn't really matter in normal desktop use.
But in datacenters loosing up to two dozen percents of performance makes lots of difference.
And now there's also TLBleed coming from Intel's implementation of SMT.
Maybe that's why Intel's looking into having no HT in 9000-serie...

That should cause big "black eye" to Intel's public "image" as maker of best CPUs.

 

[SiDeards73]

They didn't deny that they can't meet demand, but they did deny that they're outsourcing to TSMC by saying they're investing in more 14nm production lines in their own fabs.
Increased pricing is due to demand being so high, they're in a position to do that, albeit not great for consumers.

I bet increased pricing is due to them having to juggle lines about to keep up with demand and at the same time causing demand due to their capacity issue, its the perfect storm for them, they can increase prices as they slow down outbound supply due to their capacity constraints. Hence the increases we have seen creeping into their products in the past week or so... some of that will be gouging but i bet some of that is also due to supply / demand.

Also if im Intel and i have a new product about to launch and i expect it to sell well i would want a lot of units ready to go on day one to avoid the paperlaunch of Coffee Lake, these chips also need supporting chipsets, why not slow down some of the current lines to create demand, let their prices creep up, launch plenty of new product at an artificially inflated price that they themselves have managed to cause, and reap the rewards...

With AMDs Zen2 on the horizon Intel need to make hay while the sun shines so to speak.

 

[LePhuronn]

But in datacenters loosing up to two dozen percents of performance makes lots of difference.

And many big players saw this and started jumping ship. The list of EPYC customers I posted is positively eye-watering.

 

[AndreiD]

@SiDeards73 I highly doubt they're slowing down supply, it makes no sense. There's just increased demand and they can't meet it, that's all there is to it. Increased pricing always happens when demand outstrips supply (ex: DRAM pricing), it's just how the supply/demand dynamic works.

But re-reading their statement I do have to point out that I was wrong and they didn't deny that they're outsourcing to TSMC. Long term they're investing in their own production lines, but they did not clarify for short term, which means they'll most likely outsource chipsets to TSMC. They only denied outsourcing CPUs, which of course wouldn't make any sense since their 14nm is higher performance than anything comparable from TSMC.

AMD's Zen 2 is quite a bit off on the horizon, have they even taped out?

@LePhuronn Epyc uptake in data center is still sluggish: "Mercury estimates Epyc revenue was $57.66m in the second 2018 quarter vs $36m in the prior quarter."
For reference Intel's data center revenue in Q2 2018 (of which a large part is their Xeon CPUs) is $5.5 billion, growing 6% QoQ or 27% YoY. Don't expect any miracles yet.

 

[LePhuronn]

AMD's Zen 2 is quite a bit off on the horizon, have they even taped out?

Well if you use Uncle Jensen's time metrics then 6 months would qualify as "quite a bit off on the horizon" the 7nm kit is Zen 2.