Intel Set to Fall Further Behind TSMC

[Derfnofred]

I do wonder which part of the process flow is rate limiting -- most likely the litho step, although could be in the lower metal layers. In which case there's going to be a natural leveling of the field as the suppliers become more and more the rate limiter. Intel is probably extensively working with ASML and Cymer to get this process working, and will benefit pretty much everyone else.

The court of public opinion is one thing, but who also buys a chip based on the process it's on versus the performance of said chip?

 

[juanrga]

Source? Cause this sounds very made up.

Source is juanrga.

So 100% made up.

So are they having that much trouble that they can't even skip to 5 is TSMC can do it?
Seems like you might as well skip?

Intel 14nm is a pure 14nm process, whereas Globalfoundries 14nm is a hybrid node with FinFETs over a 20nm BEOL. TSMC 16nm is also a hybrid node with a 20nm BEOL. Moreover, the new 12nm and 11nm foundry nodes are rebrands of existent nodes. For instance Globafoundry 12nm is just an extension of 14nm.

https://semiengineering.com/nodes-vs-node-lets/

TSMC and Samsung 10nm nodes are closer to Intel 14nm, whereas TSMC and Globalfoundries 7nm are closer to Intel 10nm. Basic geometry parameters for 10/7nm are

TSMC 7nm:
54 nm Gate Pitch
40 nm MM Pitch
0.027µm² HD SRAM

Globalfoundries 7nm:
56 nm Gate Pitch
40 nm MM Pitch
0.0269µm² HD SRAM

Intel 10nm:
54 nm Gate Pitch
36 nm MM Pitch
0.0312µm² HD SRAM

For logic the nodes scale as

But geometric parameters don't tell all the history. Intel difficulties with 10nm node seem to be motivated by aggressive usage of new materials as Cobalt. Cobalt has different electrothermal properties and allows better scaling and potentially higher clocks than Copper.

 

[ChadD]

I would agree that Intels 10mn is comparable and perhaps even slightly better (on paper anyway) then the 7nm options out there.

Of course none of that amounts to nothing if you don't ship product. TSMC tapped out 18 customer products on their 7nm process.

I hope Intel figures it out soon. One product cycle isn't a big deal.. AMD isn't likely to ship much more then 7nm Epyc for awhile. Considering the lead Intel has had they have sometime to get this right and shipping. Having said that if AMD manages to start shipping ZEN2 7nm before Intel has their process up and running 2019/20 are going to be hard on Intels marketing budget.

That is just the ouch we know is coming for sure. There is always the dark horse A12 from Apple as well which has also now tapped out... and some early leaks tend to suggest it may well be the chip that powers the Apple move from Intel. (perhaps who knows... but the early leaks claiming 30-40% performance gains over A11 with what appears to be purposeful increased power draw are perhaps telling) I know that Intel isn't exactly making a fortune on Apple macbooks... still its potential lost business and the real threat there is if the Apple marketing machine goes to work selling people on ARM.

 

[Kinestron]

When AMD makes over a billion dollars profit in one quarter then Intel may have something to worry about. As well as AMD executed this last year, Intel is such a behemoth one of their quarterly profits made more than AMD all year(4.4 billion vs 300million). It seems like AMD is selling just about every widget they make but still can't touch Intel's profits. I wish it weren't so and feel they deserve a bigger piece of the pie but Intel's trouble with a new process is just a small bump in the road at this time. I feel like Intel can still 'coast' another couple years while they work the kinks out of 7nm and 5nm.

 

[knowom]

I agree. Intel sells 36 thread CPUs and they run up to $9000.00. Not exactly a simple thing to make. They just sat back and milked the enterprise while keeping gamers and prosumers from cannibalizing their Xeon sales by not offering cheap many core CPUs. Then came Zen and Threadripper and I'm sure they are freaking out.

Why they are just glued together...nothing to worry about.

 

[Sycraft]

I would agree that Intels 10mn is comparable and perhaps even slightly better (on paper anyway) then the 7nm options out there.

Of course none of that amounts to nothing if you don't ship product. TSMC tapped out 18 customer products on their 7nm process.

Always be careful with what companies claim though... particularly TSMC. They have a history of saying "Oh ya, this new process is totally ready, right around the corner, we are shipping soon!" only to have delays, delays, and more delays.

Basically with lithography from anyone these days, don't plan on it until you are seeing it in the hands of consumers. Too many things can crop up unexpectedly and companies love to over-promise.

 

[sleepeeg3]

Those that think intel was slacking are clueless. The PTD engineers have the hardest life in the company. They've been flogging the employees for years trying to get 10nm on track. It is just incredibly difficult and that is compounded by high turnover and little communication between teams that worked on successive generations.

Maybe they should try less flogging and more rewarding? Then their turnover would probably be lower.

 

[ChadD]

Always be careful with what companies claim though... particularly TSMC. They have a history of saying "Oh ya, this new process is totally ready, right around the corner, we are shipping soon!" only to have delays, delays, and more delays.

Basically with lithography from anyone these days, don't plan on it until you are seeing it in the hands of consumers. Too many things can crop up unexpectedly and companies love to over-promise.

They announced shipping parts like a week back or so. Apple A11 benchmarks are leaking... so those are out there for sure. As you say though the rest who knows. AMD is still saying late 2018 for Zen2 7nm, but its very likely that will be Epyc with desktop parts not shipping to 2019. We'll see I guess.

 

[Fuzzy_3D]

Maybe they should try less flogging and more rewarding? Then their turnover would probably be lower.

The flogging will continue until yields improve.

 

[N4CR]

Damage control is always funny to see.

Note that Intel's SRAM size for high-density is 0.0312um2 and for high performance, it is 0.0441um2

Most people are conveniently comparing the Intel high density number, because it looks better. The reality is even worse. For high performance, SRAM cell size is 0.0441um2 for Intel and for AMD it's 0.0353 um2, which is nearly 25% difference, instead of the 15% figure (high density, not relevant to high performance processes in question here) that is pushed around by most..

GlobalFoundries, which AMD consistently uses for its Ryzen, Polaris and Vega chips, also has a higher density than Intel's. For high-density SRAM, it is 0.0269um2 and for high-performance SRAM it is 0.0353 um2.

Funny that eh!

This difference will have a large bearing on how this process plays out. Intel has Cobalt in its' medium instead of copper, so in future spins it may provide a speed advantage, but for now, 24%+ SRAM difference is going to really help level the playing field.

All the charts I've seen fail to mention the above.

 

[Verge]

Intel's 10nm process is roughly 1.7x the raw logic density of the next densest 10nm process, albeit due to aggressive pattering techniques they also have the most complex process available to date.

https://en.wikichip.org/wiki/10_nm_lithography_process

Well to date they have no process really... but thx for explanation. Most of it is over my head, but the differences in the processes make the numeric part somewhat less meaningful.

 

[Derfnofred]

Most people are conveniently comparing the Intel high density number, because it looks better. The reality is even worse. For high performance, SRAM cell size is 0.0441um2 for Intel and for AMD it's 0.0353 um2, which is nearly 25% difference, instead of the 15% figure (high density, not relevant to high performance processes in question here) that is pushed around by most..

GF needs a lot more dead transistors in their design libraries, so logic densities outside of SRAM heavy areas is a bit worse. The GF 7nm is great at SRAM though, credit where due. Intel has definitely lost the huge process advantage it held, and now we're looking at near-parity from the big 4 (albeit Intel's process is built pretty much solely for HPC whereas everyone else looks more mobile/power scaling). What we don't have is a common chip architecture to benchmark on various processes.

 

[Fuzzy_3D]

Intel has definitely lost the huge process advantage it held, and now we're looking at near-parity from the big 4.

By what metric?

What we don't have is a common chip architecture to benchmark on various processes.

Barring that it's impossible to compare them. Some of those apples start to look a lot like oranges and bananas on closer inspection, if you know what I mean.

 

[Derfnofred]

Fuzzy, I'm not sure if you're being obtuse, but what process were the other 3 leading-edge fabs on 4 years ago? What was Intel doing? Power/Frequency/Density were all Intel's advantage. As I said, each fab is focusing on a slightly different envelope, so different chips are going to be best on different processes.

But I'm willing to trust the folks over at Chipworks and other industry experts who have a much closer ear to the ground for who's pushing processes forward. And it's historically been Intel in the bulk Si leading-edge world. Now the other 3 are right there.

 

[juanrga]

Always be careful with what companies claim though... particularly TSMC. They have a history of saying "Oh ya, this new process is totally ready, right around the corner, we are shipping soon!" only to have delays, delays, and more delays.

Basically with lithography from anyone these days, don't plan on it until you are seeing it in the hands of consumers. Too many things can crop up unexpectedly and companies love to over-promise.

To be fair TSMC's execution in last years have been very good and currently they are leaders with their 10nm node which is ahead of every other foundry node. Check the image in #42, TSMC 10nm node is only slightly denser than Intel 14nm, but is more dense.

I think your description fits better with Glofo record: 32nm delayed, 28nm delayed, 20nm canceled, first 14nm attempt (14XM) canceled, second 14nm attempt (14LPP) delayed, 12LP isn't even offered in their website, first 10nm attempt (10XM) canceled, second 10nm attempt (10LP) canceled,... we will see if their first 7nm attempt (7LP) breaks the tendency or not.

 

[juanrga]

Most people are conveniently comparing the Intel high density number, because it looks better. The reality is even worse. For high performance, SRAM cell size is 0.0441um2 for Intel and for AMD it's 0.0353 um2, which is nearly 25% difference, instead of the 15% figure (high density, not relevant to high performance processes in question here) that is pushed around by most..

There are least two reasons why analysts mention only high density SRAM: (i) TSMC, Samsung, and other foundries don't share data for high performance cells, so a comparison of foundries nodes only can be made using high density cells; (ii) high-performance cells are only used in tiny structures as register files, the large L3 caches you can find in chips use high-density SRAM; so high-density SRAM cell size is a better metric to estimate chips size on different nodes.

 

[viscountalpha]

The flogging will continue until yields improve.

I will have to use this on a friend who used to work for Intel.