Intel kills 10nm ?? oO

[kleox64]

You can only go so far though before the success rate using just multi-patterning is so low it is essentially a failure so it is effectively a deciding factor on whether a process will work (as in commercially viable) or not.

Totally depends on the design and what is being produced...

My problem with EUV is that its only a stop-gap solution, you have dozen's of other issues to worry about when you get to <5nm

 

[Rroff]

Totally depends on the design and what is being produced...

Isn't that the problem though? at that level a wide range of products are difficult to produce with traditional lithography.

My problem with EUV is that its only a stop-gap solution, you have dozen's of other issues to worry about when you get to <5nm

Sub 5nm is a whole lot of problems of its own anyhow - I imagine a lot of consumer stuff will be on refinements of 7nm for awhile.

 

[EsaT]

Intel's validation is what holds the mass production part back, what other foundries say is OK isn't OK for Intel as the validation spec is extremely aggressive... the first parts to roll off are server grade so need to qualify to the server validation spec.

Actually didn't Intel start 14nm with small very low clock dual core+iGPU Broadwell laptop chips?
Something 10nm has now failed to produce with any workable yields.
I mean they had to even disable iGPU from that paper release model.

So why would anyone think they're now trying to make server products using it?
Granted high core count server chips would need die shrink, but if even small chips haven't worked...


Though certainly agree that different nodes aren't developed in vacuum independently from each others.
Failure of earlier node is certainly going to cause extra headache for next node.

Its all very like when Intel were having issues with 90nm back in the Prescott P4 days. Back then they transitioned to 90nm but it performed worse than 130nm which Northwood was made on. The process was hot, leaky and gave poor yields.

At least 90nm node made something you could buy from shops... and without being three years behind schedule.

 

[4K8KW10]

Intel's validation is what holds the mass production part back, what other foundries say is OK isn't OK for Intel as the validation spec is extremely aggressive...

Wondering how their TIM passes, though

 

[Rroff]

Wondering how their TIM passes, though

Hey don't be knocking on Intel's magic formulate Theodent 300 TIM.

 

[DragonQ]

Actually didn't Intel start 14nm with small very low clock dual core+iGPU Broadwell laptop chips?
Something 10nm has now failed to produce with any workable yields.
I mean they had to even disable iGPU from that paper release model.

So why would anyone think they're now trying to make server products using it?
Granted high core count server chips would need die shrink, but if even small chips haven't worked...


Though certainly agree that different nodes aren't developed in vacuum independently from each others.
Failure of earlier node is certainly going to cause extra headache for next node.

14nm started with a whimper with only two desktop Broadwell chips, which were only useful for certain workloads. It was essentially the same problem Intel will have with 10nm: they couldn't get the clocks as high as the previous 22nm process at first, so they had to compensate by improving something else. In the case of Broadwell-C, it was the IGP with its 128 MiB of eDRAM (and smaller L3 cache than Haswell). It did have a slight IPC bump but couldn't reach the clocks of Haswell, which is why it was never released in a "K" version for desktop.

I suspect Icelake will be the same: lower clocks than Coffee Lake but with an IPC bump (finally). If they cannot get better overall performance I suspect they will just do a fourth revision of Skylake on 14nm+++.

 

[EsaT]

Hey don't be knocking on Intel's magic formulate Theodent 300 TIM.

Wasn't that magic formula used up bubblegums of Krzanich?

 

[Rroff]

Wasn't that magic formula used up bubblegums of Krzanich?

You have to pay the extra for the 9900K, etc. to get that feature.

 

[kleox64]

Actually didn't Intel start 14nm with small very low clock dual core+iGPU Broadwell laptop chips?
Something 10nm has now failed to produce with any workable yields.
I mean they had to even disable iGPU from that paper release model.

So why would anyone think they're now trying to make server products using it?
Granted high core count server chips would need die shrink, but if even small chips haven't worked...


Though certainly agree that different nodes aren't developed in vacuum independently from each others.
Failure of earlier node is certainly going to cause extra headache for next node.


At least 90nm node made something you could buy from shops... and without being three years behind schedule.

Server parts come out first because they provide the quickest and largest ROI, yes they are the most difficult to get working and validate however Intel can afford this approach as once you have it done it’s plain sailing with the rest of the dies.

Production actually starts with ring oscillators, SRAM, controller blocks, ALU’s, very small CPU’s etc... the data is fed back into the process and tweaked accordingly. None of this is sold or actually makes money.

 

[EsaT]

Server parts come out first because they provide the quickest and largest ROI, yes they are the most difficult to get working and validate however Intel can afford this approach as once you have it done it’s plain sailing with the rest of the dies.

Fancy thing but release dates never show server parts coming out first.
In case of 22nm only single socket Xeons (mostly identical to desktop part) came out with desktop parts.
Multi-socket Xeons for servers were year+ behind.
Exactly same for 14nm.

Also in case of architectural changes on same node, multi-socket server parts have always lagged same time behind.

You have to pay the extra for the 9900K, etc. to get that feature.

With Krzanich kicked out at summer and running low on stocks of chewed up bubblegum it figures why Intel increased price levels...

 

[kleox64]

Fancy thing but release dates never show server parts coming out first.
In case of 22nm only single socket Xeons (mostly identical to desktop part) came out with desktop parts.
Multi-socket Xeons for servers were year+ behind.
Exactly same for 14nm.

Also in case of architectural changes on same node, multi-socket server parts have always lagged same time behind.



With Krzanich kicked out at summer and running low on stocks of chewed up bubblegum it figures why Intel increased price levels...

They go to OEM’s first for testing/validation typically 6-12 months before retail.

 

[drunkenmaster]

They go to OEM’s first for testing/validation typically 6-12 months before retail.

If they aren't out, they aren't out, and desktop chips also go out for validation.

Being able to make ANY chips and being able to release to ship for revenue are entirely different things. You can make a run of chips with a 3% yield, get 50 working chips from several wafers and send them out to OEMs, that doesn't mean they can or will ship at that point.

Server only came 'first' when desktop and server used the exact same chips and there wasn't any real die size difference. IE if Intel had a P4 they used for both desktop and server, the exact same chip, why wouldn't they sell it first for server. When server chips started to be completely different and much larger, they don't come first and haven't for many generations. Not even close to coming first.

Zen 2 is going to be different simply because they are using the same chiplets and again with the same chips at the same size, the most important market money wise comes first. This would likely not be the case if the i/o die was made on 7nm and most certainly wouldn't be true if it was monolithic.

10nm has already 'launched' and the first chips on show were low yield dual core parts, 10nm dual/quad core mobile low power chips is what Intel is aiming for in late 2019, sounds like it will probably be several months into 2020 before we get desktop parts which are probably in the 120-150mm^2 range and from what we've seen of OEM leaks in past days, 14nm Cooperlake server is only hitting volume late 2019/early 2020 and icelake server is likely going to be Q3/q4 2020.

Process nodes are simple, yields are defined by defect rate and die size. With the same defect rate, which is down to the node, then yields go down exponentially with die size. <100mm^2 quad core (on 10nm, dual core on 14nm) will always have higher yields than a 100-200mm^2 die, which will always have a higher yield than 300-400mm^2 dies and so on and so on.

Early on a node when yields are worse and defect rate is higher profit on exceptionally low yield larger dies aren't actually financially viable. There is little point selling a product that both has little margin due to low yields and kills demand for existing chips but you can't produce in volume because of again, the low yields.

 

[4K8KW10]

10nm process cancelled all together?

Intel roadmap until 2021 leaked:
https://twitter.com/witeken/status/1121052220072583174

10C desktop consumer CPUs until early 2021.

 

[drunkenmaster]

We'll see, Intel are far too, well deceptive to be believed. We might see another 10nm mobile launch of chips no one really wants because they don't perform and then supposed follow up desktop chips just get cancelled, maybe Intel will actually launch follow up decent viable 10nm chips and surprise everyone, maybe they'll actually say 10nm sucks, we can't launch it, we'll be on 14nm till 2021. Intel have lied about where they are with process nodes for 6 years now, they've lied over and over and over again. They've basically lost any real credibility they have when it comes to promises made.

At this stage I'll believe they are moving to 10nm and ready for full production when they launch a mobile chip, a desktop chip following it up in say 6-8 months and that chip has what we'd deem 10nm characteristics, vastly higher density and vastly improved power efficiency compared to their 14nm chips. Until that happens pretty much anything is possible.

 

[smilingcrow]

The long shot is they jump straight to 7nm with possibly a limited 10nm release.
Considering their 10nm woes I wouldn't bet on that but considering their past form it's not a massive jump to think they could get back on track with 7nm which is being developed in parallel with lessons learnt from the 10nm fiasco.

 

[Freddie1980]

The long shot is they jump straight to 7nm with possibly a limited 10nm release.
Considering their 10nm woes I wouldn't bet on that but considering their past form it's not a massive jump to think they could get back on track with 7nm which is being developed in parallel with lessons learnt from the 10nm fiasco.

Limited in what sense? Intel were shipping out 10nm chips to partners at the end of 2018 (Canon Lake), these are only appearing in Laptops for now so maybe this will be the market for Intel's 10nm for the time being?

 

[Journey]

Limited in what sense? Intel were shipping out 10nm chips to partners at the end of 2018 (Canon Lake), these are only appearing in Laptops for now so maybe this will be the market for Intel's 10nm for the time being?

Intel shipped out a handful of 10nm CPU's that were hamstrung, just to say that they had shipped some and 10nm was moving along.

Here is a quote from Bob Swan, from last weeks Q1 '19 report - "As I shared earlier, our confidence in 10nm is also improving. In addition to the manufacturing velocity improvement I described earlier, we expect to qualify our first volume 10nm product, Ice Lake, this quarter and are increasing our 10nm volume goals for the year,”

They expect their first volume product at the end of Q4 '19 and that is the U series laptop parts, 5-15w range of CPU's, add at least another year for production ramp up on 10nm for desktop parts, and potentially longer as they move server parts to 10nm first.

 

[LePhuronn]

...as they move server parts to 10nm first.

If 10nm ever ends up being viable, I can definitely see servers getting high priority in the face of EPYC Rome absolutely tearing into market share.

 

[smilingcrow]

Limited in what sense? Intel were shipping out 10nm chips to partners at the end of 2018 (Canon Lake), these are only appearing in Laptops for now so maybe this will be the market for Intel's 10nm for the time being?

You appear to be very out of the loop on Intel's 10nm woes.
This review of the sole 10nm laptop chip available on very limited release will get you more on track:
https://www.anandtech.com/show/13405/intel-10nm-cannon-lake-and-core-i3-8121u-deep-dive-review

 

[smilingcrow]

They expect their first volume product at the end of Q4 '19 and that is the U series laptop parts, 5-15w range of CPU's, add at least another year for production ramp up on 10nm for desktop parts, and potentially longer as they move server parts to 10nm first.

If 10nm ever ends up being viable, I can definitely see servers getting high priority in the face of EPYC Rome absolutely tearing into market share.

This was reported by Anandtech last week when they announced their quarterly results:

"As for 10 nm ramp in general, Intel is only talking about producing its relatively small Ice Lake-U processors in volumes this year, which is the company’s typical way of ramping up a new node. When it comes to their larger Ice Lake-SP server parts, Intel expects to launch those 10 nm Xeon products in 2020. The company says that its Ice Lake-SP CPUs will be available in less than 12 months after its Ice Lake-U products hit the market. In fact, Intel has even advised investors to expect 10 nm Xeons to arrive “rather sooner than later” in 2020, which would imply something earlier than Q4'2020."

Speculation suggests that desktop parts are still a problem as they require high clock speeds and process performance which might well both still be lacking.
For server parts ultimate clock speed is not usually so important and as others have stated it's an area where AMD can do more damage to Intel so they will prioritise that.
The fact that there was no official mention of 10nm desktop parts speaks volumes.

The big question is how soon will 7nm come and how will Intel prioritise that?
If they prioritise server first and laptop second then maybe that is why desktop will still only get 10nm even as late as 2020/21.
Could we even see Server @7nm before desktop @10nm!
That might partly depend on the strength of their designs and how they will compete with AMD @7nm & 5nm.
Some of the staff at Intel are certainly having to sweat a lot more than they did in the decade prior to Zen launching.