Intel Considers Abandoning 18A Node for 14A Chipmaking Process

[RandallFlagg]

That’s just standard cautionary language in every 10-Q filing. Tom’s is going for the clicks. Intel is not thinking about cancelling 14A in the way implied.

Yeah but Lip also said this:

"The 14A is the process node, but clearly I will make sure that I see the internal customer, external customer, and volume commitment before I put CapEx," said Lip-Bu Tan, chief executive of Intel, during the company's conference call ..

So he is basically saying they won't continue to pursue unless it will generate enough commitment up front. That's actually a tall order.

I don't like Lip's approach, I think he's a bean counter. However, I also think it makes a lot of sense for Intel to stop hopping between nodes every 6-12 months at this stage. They had to do that to catch TSMC - and they succeeded in catching them. Remember, Intel was doing 14nm (the equivalent of TSMCs 10nm) while TSMC rolled out N7 with yield and volume for two years. Five years later and they are at a technical parity on their latest node. It's hard to make up 2 years in 5 years when your competitor is not standing still.

And for Internal Intel designs, its competitors don't necessarily get first dibs on TSMCs latest node. Brand new Nvidia 50-series GPUs are essentially 5nm tech. Zen 5 isn't much more advanced with N4. It's likely that N2 / 18A will be the top nodes for the next 2-5 years, and I suspect Nvidia / AMD will largely lean on 3nm class nodes in 2026. Meanwhile Intel could go full bore with internal products on 18A.

They need to get really good at that node, volume and yield, before hopping again. 18A and 18A-P are good for that, providing a significant advantage for internal designs over the next 12-18 months. Sure it will be much more competitive in 2027+ using 18A as their competitors switch to N2, but they have time to stop and smell the roses, so to speak.

 

[Tomorrow]

I also think it makes a lot of sense for Intel to stop hopping between nodes every 6-12 months at this stage. They had to do that to catch TSMC - and they succeeded in catching them.

They only hopped between those nodes largely on paper as none of the external customers used those and even Intel's own products were few and far between.
It dont even think they matched TSMC yet. They say they do, but talk is cheap. Lets see actual products and their performance.
TSMC has proven track record with their nodes. If they say they will make 40k wafers of N2 in 2H 2025 i believe them.
I dont believe anything Intel says until i see actual products made on 18A.

Five years later and they are at a technical parity on their latest node. It's hard to make up 2 years in 5 years when your competitor is not standing still.

On paper yes, but so was 20A until it was canned.

Brand new Nvidia 50-series GPUs are essentially 5nm tech.

That's Nvidia being lazy, but still asking top dollar.

Zen 5 isn't much more advanced with N4.

Zen 5 launched in 2024, August. It's nearly a year old product at this point. Zen 6 will use N2 in 2H 2026/1H 2027.

It's likely that N2 / 18A will be the top nodes for the next 2-5 years, and I suspect Nvidia / AMD will largely lean on 3nm class nodes in 2026.

Mix of 2nm and 3nm depending on the segment. CPU's and HPC GPU's will likely use 2nm while consumer GPU's will likely use 3nm.

Meanwhile Intel could go full bore with internal products on 18A.

Not many of them if Celestial is made by TSMC and the graphics tile in Nova Lake is also made by TSMC.
Only Nova Lake's compute tile and Panther Lake are likely 18A candidates. I doubt any external customers will use 18A as there's nothing announced. By this time there would be something as these things are planned years ahead.

Even that would be progress considering Arrow Lake was entirely made by TSMC excluding the interposer made on old 14nm Intel 16 node.

 

[RandallFlagg]

They only hopped between those nodes largely on paper as none of the external customers used those and even Intel's own products were few and far between.
It dont even think they matched TSMC yet. They say they do, but talk is cheap. Lets see actual products and their performance.
TSMC has proven track record with their nodes. If they say they will make 40k wafers of N2 in 2H 2025 i believe them.
I dont believe anything Intel says until i see actual products made on 18A.

No, they had the nodes up and working, and used them. They just didn't use them for extreme volume relatively speaking. I say relatively, because they've shipped millions of Intel 4 and Intel 3 chips.

On paper yes, but so was 20A until it was canned.


That's Nvidia being lazy, but still asking top dollar.

And still making the most power efficient and fastest GPUs. I'm not so sure who is lazy there..

Zen 5 launched in 2024, August. It's nearly a year old product at this point. Zen 6 will use N2 in 2H 2026/1H 2027.

Like I said, 18A is a strategic advantage for 12-18M. Possibly longer, but at least that amount of time.

Also, go look at the sources for Zen 6 on N2.

ComputerBase, which referenced a forum post at AnandTech...

Mix of 2nm and 3nm depending on the segment. CPU's and HPC GPU's will likely use 2nm while consumer GPU's will likely use 3nm.

I bet you don't see N2 until 2027. N3 is expected.

Not many of them if Celestial is made by TSMC and the graphics tile in Nova Lake is also made by TSMC.
Only Nova Lake's compute tile and Panther Lake are likely 18A candidates.

You are talking about how they hopped, as I noted, without fully developing the Intel 4 and Intel 3 nodes.

They were just doing the bare minimum to get to the goal with 18A. That's the only way they could catch up. I already made note of this. They never fully developed for mass production any of those nodes because they were just milestones on their way to 18A.

I doubt any external customers will use 18A as there's nothing announced. By this time there would be something as these things are planned years ahead.

Intel has already announced Microsoft, Amazon, and the DoD as customers.

Even that would be progress considering Arrow Lake was entirely made by TSMC excluding the interposer made on old 14nm Intel 16 node.

Intel 16 is Intel's old 22nm optimized for foundry work, with FinFet added. MediaTek and UMC both use this node, and Intel + UMC are partnering to make a 12nm node.

 

[Tomorrow]

And still making the most power efficient and fastest GPUs. I'm not so sure who is lazy there.

Really?

9070 XT is 304W TDP. RTX 5080 is 360W TDP.
That's TDP difference of ~18% and RTX 5080 is ~20% faster.

If there is efficiency advantage it is so miniscule it might as well be unnoticeable.
For someone playing at 4K that 20% means going from 50fps to 60fps. While being a whopping 2% more efficient...

Frankly Nvidia should be ashamed at phoning home Blackwell. Using expensive G7 memory and still not being noticeably ahead.
And they even managed to screw up drivers somehow.

Also, go look at the sources for Zen 6 on N2.

ComputerBase, which referenced a forum post at AnandTech...

As if sources for 18A's supposed success are more reliable...

I bet you don't see N2 until 2027. N3 is expected.

We will see.

Intel has already announced Microsoft, Amazon, and the DoD as customers.

Like i said before - until there are actual products we wont know if it's any good. Cost, yields, volume etc.

 

[RandallFlagg]

Really?

9070 XT is 304W TDP. RTX 5080 is 360W TDP.
That's TDP difference of ~18% and RTX 5080 is ~20% faster.

The 9070 XT is not in the same league as the 5080. At best, it is a 5070 Ti competitor.

You are obviously ignorant of another well known fact : All of these GPUs get less and less efficient as you move up the performance stack. Comparing two different performance tier GPUs from different vendors for power efficiency is very misleading and disingenuous.

AMD does not have a 5080 competitor. At. All.

Power:

Performance:

If there is efficiency advantage it is so miniscule it might as well be unnoticeable.
For someone playing at 4K that 20% means going from 50fps to 60fps. While being a whopping 2% more efficient...

Sure, but here you have a 5nm chip whipping a 3nm chip.

In power. And performance.

Frankly Nvidia should be ashamed at phoning home Blackwell. Using expensive G7 memory and still not being noticeably ahead.
And they even managed to screw up drivers somehow.

I just replaced an AMD GPU with Nvidia. Not because of small power / performance differences, because of better performance : price ratio, and much better support for AI and DLSS 4.

 

[Visible Noise]

So he is basically saying they won't continue to pursue unless it will generate enough commitment up front. That's actually a tall order.

CapEx is capital expenditure. Development is expense, not capital.

What he is saying is he’s not going on a buying binge for tooling (capital expenditure) until the node is ready for production and has enough volume to bring up new fab lines (which costs billions of dollars) Intel always does first production in either Oregon or Arizona development fab lines. Once volume is up and the process is smooth then they do their “copy everywhere” strategy of cloning that fab line out to the other factories.

In simpler terms, Intel is conserving cash, not canceling their future. At least not yet.

 

[Tomorrow]

The 9070 XT is not in the same league as the 5080. At best, it is a 5070 Ti competitor.

You are obviously ignorant of another well known fact : All of these GPUs get less and less efficient as you move up the performance stack. Comparing two different performance tier GPUs from different vendors for power efficiency is very misleading and disingenuous.

Fine. 5070 Ti has 300W TDP. 9070 XT is 304W. 1.3% difference in power and 4% difference in performance. The end result is not much different from 5080 and 9070 XT efficiency comparison.

Again. I fail to see this supposed power efficiency advantage Blackwell supposedly has over RDNA4.
The 8% advantage in the image you provided could very well be caused by different AIB models having different characteristics.

Im going by the official numbers in TPU's GPU database.

Sure, but here you have a 5nm chip whipping a 3nm chip.

In power. And performance.

Since when is 9070 XT a 3nm chip? They're both 5nm class chips. The differences between 4N and N4P are pretty miniscule.
Also 5070 Ti has 39% higher memory bandwidth thanks to G7. Yet is only 4% faster. Like i said - embarrassing.
Recent testing even puts that 4% into question as some updates (other than just drivers) seem to have improved RDNA4 performance:

I just replaced an AMD GPU with Nvidia. Not because of small power / performance differences, because of better performance : price ratio, and much better support for AI and DLSS 4.

And i replaced Nvidia with AMD.