Intel 20A Node Cancelled for Foundry Customers, "Arrow Lake" Mainly Manufactured Externally

[tpuuser256]

Lithographic nodes, have reached a size beyond which reduction will primarily lead to an increase in manufacturing defects. The right way is to stop the race to scale down and work on architectural improvements to the compute lines. Gateway width, more gateways, more and smart gates, more efficient predictors. Bigger and smart caches.

You should apply to be CTO of NVidia

 

[londiste]

The think is that their 7nm (Intel 4) is garbage, it has 120 MTr/mm2, for comparison, their 10nm (intel 7) is 100 MTr/mm2. No wonder it's marginal better than their polished 10nm. Now lets see the TSMC 3nm, that will be used for Arrow and Lunar Lake - 200-220 MTr/mm2 and their mythical 18a won't be more than 160 MTr/mm2, 180 at most and it will be ready for 2026. They will release Arrow Lake in the end of 2024, refresh it with the same TSMC node in 2025 (Arrow Lake refresh) and when they are ready to release whatever is real upgrade over Arrow Lake, it will be in 2026 and how you see them to downgrade from 220 MTr/mm2 to their miserable 160 MTr/mm2? They won't, they will go again for the next TSMC node. Their fabs are so behind, that its no longer cost effective to build their own CPUs here, because this will hurt the performance and the competitiveness

You are comparing different things here. Intel 4 and TSMC N3 are roughly in the same ballpark for high performance cells. TSMC N3 has the high-density variation has the higher density but obviously with performance implications. If they have a problem it is somewhere in manufacturing.

 

[phints]

So is Lunar Lake and Arrow Lake both TSMC N4 like the rest of the top chips use?

Better question is, Arrow Lake when?

 

[phanbuey]

The think is that their 7nm (Intel 4) is garbage, it has 120 MTr/mm2, for comparison, their 10nm (intel 7) is 100 MTr/mm2. No wonder it's marginal better than their polished 10nm. Now lets see the TSMC 3nm, that will be used for Arrow and Lunar Lake - 200-220 MTr/mm2 and their mythical 18a won't be more than 160 MTr/mm2, 180 at most and it will be ready for 2026. They will release Arrow Lake in the end of 2024, refresh it with the same TSMC node in 2025 (Arrow Lake refresh) and when they are ready to release whatever is real upgrade over Arrow Lake, it will be in 2026 and how you see them to downgrade from 220 MTr/mm2 to their miserable 160 MTr/mm2? They won't, they will go again for the next TSMC node. Their fabs are so behind, that its no longer cost effective to build their own CPUs here, because this will hurt the performance and the competitiveness

Right but I doubt the USA will care to let the last even remotely competitive fab fall over and rely solely on TSMC.

I can see tax breaks and incentives in the future to pump these fabs.

 

[Daven]

If Arrow Lake is still coming out in a month, then it was always all TSMC. No way you can switch nodes last minute without massive delays.

Arrow Lake P cores on Intel 20A was another lie.

 

[Eternit]

The think is that their 7nm (Intel 4) is garbage, it has 120 MTr/mm2, for comparison, their 10nm (intel 7) is 100 MTr/mm2. No wonder it's marginal better than their polished 10nm. Now lets see the TSMC 3nm, that will be used for Arrow and Lunar Lake - 200-220 MTr/mm2 and their mythical 18a won't be more than 160 MTr/mm2, 180 at most and it will be ready for 2026. They will release Arrow Lake in the end of 2024, refresh it with the same TSMC node in 2025 (Arrow Lake refresh) and when they are ready to release whatever is real upgrade over Arrow Lake, it will be in 2026 and how you see them to downgrade from 220 MTr/mm2 to their miserable 160 MTr/mm2? They won't, they will go again for the next TSMC node. Their fabs are so behind, that its no longer cost effective to build their own CPUs here, because this will hurt the performance and the competitiveness

Yes. But if manufacturing in their fabs their own chips is not viable and there are not a lot of contracting customers what is the point of spending bilions on the next nodes.

 

[remixedcat]

It is not his fault, but shareholders will demand someones head and CEO is natural choice. He hasn't caused these problems, but he was hired to fix Intel and he hasn't.

he's praying on twitter but God hasn't answered his prayers yet. He should have followed God's lead and became a preacher but he's the CEO of a CPU company instead

 

[londiste]

If Arrow Lake is still coming out in a month, then it was always all TSMC. No way you can switch nodes last minute without massive delays.

Arrow Lake P cores on Intel 20A was another lie.

Depends on what that "primarily" means. In terms of chiplet count and area it has always been mostly from TSMC. CPU chiplet is quite small as well.
This announcement is from the Foundry group and saying 20A is cancelled for Foundry customers does not necessarily mean Intel themselves does not use it.

 

[usiname]

Yes. But if manufacturing in their fabs their own chips is not viable and there are not a lot of contracting customers what is the point of spending bilions on the next nodes.

Its just too late to leave the semiconductor business, I don't think they can leave now with all taxpayers funds, they burned for broken nodes

 

[TumbleGeorge]

Its just too late to leave the semiconductor business, I don't think they can leave now with all taxpayers funds, they burned for broken nodes

They might start making silicone stuff for plastic surgery.

 

[kondamin]

They just stack all the cash the US has been pumping into them and burn it Joker style? I don’t think I’ve heard a single positive thing about Intel’s foundries for quite some time now.

They have been building fabs with that money, those fabs aren't online yet.

it's the main reason I think it's very likely that if sh*t really hits the fan the government will bail them out and in time they will be OK.
(As those fabs without intel it self as the main customer is pointless)

 

[N/A]

So is Lunar Lake and Arrow Lake both TSMC N4 like the rest of the top chips use?

Better question is, Arrow Lake when?

N3B and this is the best node for the time being, 56% better than N4 in logic density. SRAM cell on the other hand has reached a wall, not shrinking. N2 is only 16% more dense. A16 another 10%.

Arrow lake just around the corner, but expensive.

 

[Daven]

Depends on what that "primarily" means. In terms of chiplet count and area it has always been mostly from TSMC. CPU chiplet is quite small as well.
This announcement is from the Foundry group and saying 20A is cancelled for Foundry customers does not necessarily mean Intel themselves does not use it.

The article specifically mentions Arrow Lake on external processes and just packaged by IFS.

 

[R0H1T]

They might start making silicone stuff for plastic surgery.

Nice pivot in the age of Tiktok

 

[Eternit]

They have been building fabs with that money, those fabs aren't online yet.

it's the main reason I think it's very likely that if sh*t really hits the fan the government will bail them out and in time they will be OK.
(As those fabs without intel it self as the main customer is pointless)

So they will be pointless as Intel will be TSMC customer.

 

[ThomasK]

Guess TSMC is running to the bank with Intel's money.

 

[RandallFlagg]

It wasn't making much sense to be developing two almost identical nodes at the same time, expect if Intel wasn't sure that 18A will succeed and had 20A as a back up plan. Probably 18A is doing well enough for them to not need to keep throwing money at 20A node.
That's the good scenario.

The bad scenario takes in consideration the fact, Broadcom not being enthusiastic about Intel's manufacturing in it's current phase and also seeing Intel "relying on external partners for Arrow Lake production" instead of moving it to it's own 18A manufacturing node. That probably means that Intel, not just Broadcom, is still unsure about 18A quality or scheduling. Until Intel comes out with a mass producing product at 18A, their manufacturing will continue being a huge question mark, years behind TSMC and even Samsung.

That isn't the way it works, though people who have only looked at node names in a quick and cursory way wouldn't know.

Keep in mind, node names are marketing gimmicks.

TSMC’s 7nm, 5nm, and 3nm “are just numbers… it doesn’t matter what the number is”

"These numbers are just numbers. They're like models in a car - it’s like BMW 5-series or Mazda 6."

www.pcgamesn.com

20A and 18A are essentially the same node, but 18A is mostly a newer set of 'libraries' for development combined with smaller what I would call 'trace' lines. Libraries in this context refers to how components are arranged.

20A and 18A have a similar relationship to each as N3, N3E, N3B have to each other, or a better analogy is probably N5 vs N4.

The same is true of Intel 3 and Intel 4.

Also of note, 20A was specifically just for internal use by Intel. The same is true of Intel 4, it was just for internal designs.

18A and Intel 3 are IFS nodes, meaning for use both internally and externally.

 

[tpuuser256]

They have been building fabs with that money, those fabs aren't online yet.

it's the main reason I think it's very likely that if sh*t really hits the fan the government will bail them out and in time they will be OK.
(As those fabs without intel it self as the main customer is pointless)

If I had 500k free, they would go to INTC

 

[qcmadness]

You are comparing different things here. Intel 4 and TSMC N3 are roughly in the same ballpark for high performance cells. TSMC N3 has the high-density variation has the higher density but obviously with performance implications. If they have a problem it is somewhere in manufacturing.

The real difference is one is in very high volume manufacturing and one is on paper.

 

[Tek-Check]

What would shareholders say to slogan "5 nodes in 4 years" when one of those nodes is cancelled? Another law suit?

 

[RandallFlagg]

The real difference is one is in very high volume manufacturing and one is on paper.

Well that's one way you can demonstrate your complete ignorance.

Meteor Lake is using Intel 4.

Intel Core Meteor Lake (Intel4) I/O Die Digital Floorplan Analysis | TechInsights

Intel Core Meteor Lake (Intel4) I/O Die Digital Floorplan Analysis

www.techinsights.com

Moreover, Granite Rapids and Sierra Forest are on Intel 3.

Intel Launches 144-core 'Sierra Forrest' Xeon 6 CPUs, Granite Rapids Follows in Q3

Pumping up the core counts.

www.tomshardware.com

 

[Eternit]

Well that's one way you can demonstrate your complete ignorance.

Meteor Lake is using Intel 4.

Intel Core Meteor Lake (Intel4) I/O Die Digital Floorplan Analysis | TechInsights

Intel Core Meteor Lake (Intel4) I/O Die Digital Floorplan Analysis

www.techinsights.com

Moreover, Granite Rapids and Sierra Forest are on Intel 3.

Intel Launches 144-core 'Sierra Forrest' Xeon 6 CPUs, Granite Rapids Follows in Q3

Pumping up the core counts.

www.tomshardware.com

View attachment 362181

Meteor Lake is not produced in high volumes and it only has compute tile made on Intel 4. Granite Rapids is not produced in high quantities either and while they have large tiles, they run at rather low frequency and there is a lot of versions with more or less cores so they can do binning. The are no other plans to use Intel 4 and 3. I think Meteor Lake will be quickly replaced by Lunar Lake. These nodes will never return what Intel invested in them.

 

[Visible Noise]

I thought everyone knew 20A is an internal dev node and 18A is the foundry node???

 

[R0H1T]

Everyone? You mean no one informed Intel

Spoiler

 

[Visible Noise]

Everyone? You mean no one informed Intel

Spoiler

Where does that say foundry in the 20A section?

Here’s Intel saying 18A is the foundry node back in February, which agrees with your slide:

Intel Foundry Services Head Stu Pann explains company's plan to build Arm chips, move more manufacturing to the U.S.

Intel poised to compete with TSMC

www.tomshardware.com