News Intel Says Process Tech to Lag Competitors Until Late 2021, Will Regain Leadership with 5nm

[Admin]

Intel disclosed that it won't be able to match competing process tech until 2021, and process leadership won't come until the 5nm node.

Intel Says Process Tech to Lag Competitors Until Late 2021, Will Regain Leadership with 5nm : Read more

 

[jimmysmitty]

To me the biggest thing Intel needs to worry about is a new uArch and utilizing technologies like Foveros to implement designs like AMD is doing, or better yet like their old M2M designed Core 2 series. That is where the advantage lies.

AMD pulling the I/O off the cores allows them to create cores that use less power overall, of course the 7nm is not hurting at all, and run cooler while also allowing for higher core counts. However the down side is this does add in latency which does affect performance in some areas. Foveros looks like it should help to mitigate that with a much closer connection and allow for Intel to have similar benefits.

But who am I to say. I am just a lowly nerd on some forum.

 

[Giroro]

If this is what George Davis thinks the "10nm era" looks like, then maybe Intel needs a CFO with higher standards.
If George Davis thinks letting your competitors beat your 5nm to market by 2 years is "process leadership", then... I don't even know where to start with that one. Maybe Intel should add "ability to read a calendar" to their job requirements for CFO?
Does intel want their hypothetical first 5nm products to outperform Samsung's 4nm+ or TSMC's 5nm++? Probably. But what an executive wants and what Intel's struggling engineering department can eventually produce are two very different things.

 

[JamesSneed]

To me the biggest thing Intel needs to worry about is a new uArch and utilizing technologies like Foveros to implement designs like AMD is doing, or better yet like their old M2M designed Core 2 series. That is where the advantage lies.

AMD pulling the I/O off the cores allows them to create cores that use less power overall, of course the 7nm is not hurting at all, and run cooler while also allowing for higher core counts. However the down side is this does add in latency which does affect performance in some areas. Foveros looks like it should help to mitigate that with a much closer connection and allow for Intel to have similar benefits.

But who am I to say. I am just a lowly nerd on some forum.

I tend to agree. Albeit Intel also needs to get to 7nm using EUV on time as well. Foveros is a much more elegant solution than AMD's but then AMD's has been out a while now. I have a feeling AMD's Zen3 will reduce much of the inherent latencies.

 

[jimmysmitty]

I tend to agree. Albeit Intel also needs to get to 7nm using EUV on time as well. Foveros is a much more elegant solution than AMD's but then AMD's has been out a while now. I have a feeling AMD's Zen3 will reduce much of the inherent latencies.

Yes 7nm but they need to move on from Core and push the new Sunny Cove or later designs in more than just laptops.

Most of the latencies come from the distance from the I/O die to the cores. Foveros has almost none. I am sure though that AMD is working on some sort of similar design.

 

[spongiemaster]

If this is what George Davis thinks the "10nm era" looks like, then maybe Intel needs a CFO with higher standards.
If George Davis thinks letting your competitors beat your 5nm to market by 2 years is "process leadership", then... I don't even know where to start with that one. Maybe Intel should add "ability to read a calendar" to their job requirements for CFO?
Does intel want their hypothetical first 5nm products to outperform Samsung's 4nm+ or TSMC's 5nm++? Probably. But what an executive wants and what Intel's struggling engineering department can eventually produce are two very different things.

You can't compare node labels between companies. By TSMC's own admission, they no longer have any real world meaning to them. Intel uses the most conservative labels in the industry which is well known. Intel's 7nm will be equivalent to the competition's 5nm. So it doesn't matter that what Intel will call its 5nm node will hit the market 2 years after what TSMC calls its 5nm because they aren't comparably dense nodes. Intel doesn't need to "want" their 5nm product to beat the competition's 5nm, because we already know it will. It should be equivalent to what TSMC will call its 3nm node.

 

[ta152h]

This is an interesting article, because you rarely hear Intel disparage their own products and fabrication nodes, but clearly that's what happened here.

Even so, 14nm++ is better than anything TSMC has in certain metrics; it's doubtful Intel's clock speed advantage is entirely architectural considering how much it keeps going up with each refresh of 14nm. The main problem is, the architecture is very old. Plus, it's been degraded by roughly 15 % or so by security mitigations. Yet, AMD still can't match the clock speed or single threaded performance, so it's not all bad.

10nm has the new architecture, and that alone helps a lot, even if the node itself isn't particularly strong. Roughly 18% IPC improvement forgives a lot of clock speed, and Intel's charts show later 10nm nodes having better performance than 14nm++, so they should have some decent products available.

Of course, they keep getting blasted by security holes, so if that keeps up, it might not go too well.

 

[InvalidError]

Yes 7nm but they need to move on from Core and push the new Sunny Cove or later designs in more than just laptops.

Except they can't do it until they have the new process required to maintain or increase clock frequencies on a more complex chip - there is little point in increasing IPC by 20% if it comes at the expense of 20% lower clocks since you end up with more expensive chips that don't perform any better.

Most of the latencies come from the distance from the I/O die to the cores. Foveros has almost none.

The distance is not the real problem: 10mm at 200e6 m/s is only 50ps. The real kicker is shifting the data through the chip-to-chip protocol which takes several chip-to-chip bus clock ticks in each direction. That's where the bulk of Zen 2's ~20ns chiplet-RAM latency penalty comes from. Wire distance accounts for less than 1/10th of a fabric clock cycle, effectively immaterial.

 

[Giroro]

You can't compare node labels between companies. By TSMC's own admission, they no longer have any real world meaning to them. Intel uses the most conservative labels in the industry which is well known. Intel's 7nm will be equivalent to the competition's 5nm. So it doesn't matter that what Intel will call its 5nm node will hit the market 2 years after what TSMC calls its 5nm because they aren't comparably dense nodes. Intel doesn't need to "want" their 5nm product to beat the competition's 5nm, because we already know it will. It should be equivalent to what TSMC will call its 3nm node.

Whether or not Intel's 5nm will beat their competition in density, and whether or not they can even reliably produce it remains to be seen. It is a purely hypothetical product right now.
If Intel were far enough into their 5nm development to produce it and thoroughly evaluate their technology against the competition, then why are they still having so many problems bringing 10nm to market?
Intel's 10nm is denser than TSMC's 1st gen 7nm, but there isn't any any reason to assume the same of future products, especially when their 10nm strategy isn't working out very well for them.

The names of the products might be influenced by marketing buzz, but marketing matters too.

 

[jimmysmitty]

Whether or not Intel's 5nm will beat their competition in density, and whether or not they can even reliably produce it remains to be seen. It is a purely hypothetical product right now.
If Intel were far enough into their 5nm development to produce it and thoroughly evaluate their technology against the competition, then why are they still having so many problems bringing 10nm to market?
Intel's 10nm is denser than TSMC's 1st gen 7nm, but there isn't any any reason to assume the same of future products, especially when their 10nm strategy isn't working out very well for them.

The names of the products might be influenced by marketing buzz, but marketing matters too.

I don't doubt that Intel will still be ahead in density.

As for marketing, outside of enthusiast circles and the tech world itself the mass majority do no know what is meant by process size. And if you want to talk marketing the majority of people know Intel more than they do AMD. Hell even newer Dell commercials still have the Intel logo with them.

 

[InvalidError]

Whether or not Intel's 5nm will beat their competition in density, and whether or not they can even reliably produce it remains to be seen. It is a purely hypothetical product right now.
If Intel were far enough into their 5nm development to produce it and thoroughly evaluate their technology against the competition, then why are they still having so many problems bringing 10nm to market?

Intel began construction of new fabs for 7nm and beyond only a year ago and it takes about three years from breaking ground to first production silicon, so there won't be any meaningful 5nm production for another two years no matter how far along Intel may be on the process R&D side of things. As for why Intel is struggling with 10nm, that's simply from pushing DUV to its practical limits.

Stepping up to EUV will eliminate many of the complications Intel ran into with 10nm DUV and it is also very likely that Intel will be able to apply what it has learned from pushing DUV to the extremes of what it can achieve to EUV, in which case it may be able to leapfrog everyone else on EUV once its EUV fabs are ready.

Doing 10nm on DUV is a bit like trying to paint a fine portrait using only 10" rollers and EUV is an upgrade to 1" brushes. You still need clever tricks to make it work, just not as many.

 

[jimmysmitty]

Intel began construction of new fabs for 7nm and beyond only a year ago and it takes about three years from breaking ground to first production silicon, so there won't be any meaningful 5nm production for another two years no matter how far along Intel may be on the process R&D side of things. As for why Intel is struggling with 10nm, that's simply from pushing DUV to its practical limits.

Stepping up to EUV will eliminate many of the complications Intel ran into with 10nm DUV and it is also very likely that Intel will be able to apply what it has learned from pushing DUV to the extremes of what it can achieve to EUV, in which case it may be able to leapfrog everyone else on EUV once its EUV fabs are ready.

Doing 10nm on DUV is a bit like trying to paint a fine portrait using only 10" rollers and EUV is an upgrade to 1" brushes. You still need clever tricks to make it work, just not as many.

FAB 42 never entered production. It sat there doing nothing. They are fitting it for 7nm. It will be one of the first ones for it.

D1X is getting an expansion as is the facility in Israel. Those will probably take more time but I imagine FAB 42 will be able to ramp up 7nm faster since they do not have to build anything.

 

[InvalidError]

Those will probably take more time but I imagine FAB 42 will be able to ramp up 7nm faster since they do not have to build anything.

Even when you have existing buildings, they may still require a complete overhaul of their vibration dampening systems to provide sufficiently stable slabs for optical equipment handling 15X smaller wavelength. I suspect there is a whole lot more going on than just lithography equipment swaps.

 

[jimmysmitty]

Even when you have existing buildings, they may still require a complete overhaul of their vibration dampening systems to provide sufficiently stable slabs for optical equipment handling 15X smaller wavelength. I suspect there is a whole lot more going on than just lithography equipment swaps.

The only thing I have seen is a few months ago they did have some heavy duty cranes but thats to be expected since they should be moving out a lot of larger equipment.

Still FAB 42 should be a faster ramp up than a brand new building like D1X seems to be getting as a building adds quite a bit of time to it.

Next time I see a guy I know whop works for Intel I will ask. I am sure he wont be able to tell me much but I haven't seen as much heavy equipment as when they were first building it.

 

[spongiemaster]

Whether or not Intel's 5nm will beat their competition in density, and whether or not they can even reliably produce it remains to be seen. It is a purely hypothetical product right now.

Not really, no one here is projecting performance, but density is easy to project in Intel's favor based on company statements.

Intel's currently selling 10nm CPU's have a calculated density of 101 million transistors/mm2. TSMC's 7nm has a calculated density of 91MTr/mm2. TSMC claims that 5nm will be 80% more dense than its 7nm which lands it at about 165 MTr/mm2. Bob Swan claimed in July of last year Intel's 7nm will be exactly 2x more dense than its 10nm. At 202MTr/mm2, Intel's 7nm is going to be denser than TSMC's 5nm. Based on that, there is no credible argument to be made that Intel's 5nm won't be denser than TSMC's 5nm.

 

[InvalidError]

Based on that, there is no credible argument to be made that Intel's 5nm won't be denser than TSMC's 5nm.

There is at least one major reason: Intel has been consistently failing to deliver on 10nm for three years, skepticism until Intel's 7nm and 5nm are actually online and verified is justified.

 

[spongiemaster]

There is at least one major reason: Intel has been consistently failing to deliver on 10nm for three years, skepticism until Intel's 7nm and 5nm are actually online and verified is justified.

That's only a credible argument if you truly believe there is a realistic scenario where Intel never makes it to 5nm and just throws in the towel and closes up shop. As I pointed out in my last post, Intel's 7nm will be denser than TSMC's 5nm, so you would have to believe that Intel won't even make it to 7nm. I don't consider that a credible argument. Intel will make it to 7nm, thus having a denser node than TSMC's 5nm node.

 

[TerryLaze]

Except they can't do it until they have the new process required to maintain or increase clock frequencies on a more complex chip - there is little point in increasing IPC by 20% if it comes at the expense of 20% lower clocks since you end up with more expensive chips that don't perform any better.

Sure they could,if intel's only goal where to one up AMD they could easily make a CPU core with more execution units that would beat ryzen in everything.
The 3900x only hits ~3.6Ghz all core when running anything demanding,it's pretty much impossible that intel couldn't hit the same clocks,20% below 5Ghz is 4Ghz and 3.6 is another 10% below that.
The only reason intel doesn't do that is that they don't have to destroy their sequential speed to keep selling everything they can produce.

View: https://youtu.be/3LesYlfhv3o?t=405

 

[InvalidError]

Sure they could,if intel's only goal where to one up AMD they could easily make a CPU core with more execution units that would beat ryzen in everything.

If Intel could "easily" make CPUs with more cores on 10nm, it would be shifting its entire lineup to 10nm to get more dies per wafer and keep TDP in check. Instead, Intel is skipping straight to 7/5nm for mainstream. That's a pretty clear sign that Intel has given up on scaling 10nm beyond what it needs to support its limited 10nm portfolio.

An even more telling sign of how bad things are with 10nm is Intel building more 14nm production lines instead of going all-in on scaling up 10nm production and migrate more products to it as production scales up. That's basically an admission that 10nm has hit brick walls that may not be overcome within 10nm's useful life and make it non-viable for low-margin high-volume mainstream parts.

 

[st379]

I don't see Intel leading in the near future (3 years).
7 nm EUV confirmed this year with ryzen and Nvidia gpu.
Last article I read 5 nm EUV at 80% yield. Safe to say it will be coming with ryzen and maybe new gpu in 2021.
3 nm and 2 nm under development.
And while TSMC already have 7 and 5 nm Intel "claims" a lot of thing about 7 nm.

Last month, Intel say cpu on 7 nm will be coming in q1 2022. Lol it is almost 2 years from now.

 

[jimmysmitty]

I don't see Intel leading in the near future (3 years).
7 nm EUV confirmed this year with ryzen and Nvidia gpu.
Last article I read 5 nm EUV at 80% yield. Safe to say it will be coming with ryzen and maybe new gpu in 2021.
3 nm and 2 nm under development.
And while TSMC already have 7 and 5 nm Intel "claims" a lot of thing about 7 nm.

Last month, Intel say cpu on 7 nm will be coming in q1 2022. Lol it is almost 2 years from now.

That 80% yield is for a test chip thats not even a modern CPU. Typically test chips are on SRAM.

https://www.anandtech.com/show/15219/early-tsmc-5nm-test-chip-yields-80-hvm-coming-in-h1-2020

Its just a marketing headline. They do not have 80% yields on a high powered processor.

Also remember the naming means nothing now. Intel could easily just rename its 10nm to 7nm as their 10nm is slightly denser than TSMCs 7nm.

 

[st379]

I understand that it is supposed to launch on smartphones very soon. I know it is different but 5 nm for smartphones while i don't hear anything from Intel on 7 nm. Maybe q1 2022.
I think that TSMC will not encounter problems in 5nm.
Intel 7 nm will probably be against TSMC 5 nm, with no delays from Intel.
We are still talking about 2022 for Intel. 2 years from now.

 

[jimmysmitty]

I understand that it is supposed to launch on smartphones very soon. I know it is different but 5 nm for smartphones while i don't hear anything from Intel on 7 nm. Maybe q1 2022.
I think that TSMC will not encounter problems in 5nm.
Intel 7 nm will probably be against TSMC 5 nm, with no delays from Intel.
We are still talking about 2022 for Intel. 2 years from now.

Low power chips. Cannot compare.

Even so Intels 7nm, if specs are on track, will probably be as dense or more dense than TSMCs 5nm.

 

[st379]

Low power chips. Cannot compare.

Even so Intels 7nm, if specs are on track, will probably be as dense or more dense than TSMCs 5nm.

Probably the same. No leadership, and 3nm is on track too. I don't see any leadership and we are talking on 2022 with no delays. The days of 28nm vs 14nm+++ are over.
The same nm more or less and who has the better architecture.
Like I said 2 years from now it is a lot of time and a lot of things can change.

 

[TerryLaze]

If Intel could "easily" make CPUs with more cores on 10nm, it would be shifting its entire lineup to 10nm to get more dies per wafer and keep TDP in check. Instead, Intel is skipping straight to 7/5nm for mainstream. That's a pretty clear sign that Intel has given up on scaling 10nm beyond what it needs to support its limited 10nm portfolio.

I did not say more cores,I said more execution units per core,like exactly what they did with the 10th gen mobile CPUs if they can make those run at ~3.6Ghz all core and match the core count of ryzen they would have a better product than ryzen but with the same downside as well it would be very slow in running sequential code.

An even more telling sign of how bad things are with 10nm is Intel building more 14nm production lines instead of going all-in on scaling up 10nm production and migrate more products to it as production scales up. That's basically an admission that 10nm has hit brick walls that may not be overcome within 10nm's useful life and make it non-viable for low-margin high-volume mainstream parts.

Yes,why ever would intel milk 14nm for everything they can when they could burn through nm as if there were limitless potential to go smaller and smaller...
Everybody else will hit a brick wall in not being able to go smaller while intel will still be making all the money in the world on 14nm,well maybe they will be on 10nm+++++++ by then.