Intel Unveils Panther Lake Architecture: First AI PC Platform Built on 18A

[tooLongInEDA]

Intel was able to do 5 nodes in 4 years because the lead fab has the mission to lead, seed and the seeds transfer to another fab. It used to be every other year a new node, before all the difficulty starting in the 2010 until present. The Ronler Acres Moore Park complex is a grand scale. There is probably a cheaper way to do R&D but that is the Intel way. It worked better when fab lite wasn't a viable option as it is now.

On the topic of taking fab lite off the table for Intel, so they focus on the old tick tock, now that Intel's capital structure is different, with Trump and SoftBank and Nvidia, perhaps they can sack the board members who still want to see fab lite happen? Make a statement that fab lite is dead?

I do hope the official history (if one ever exists) doesn't record Pat Gelsinger's claim of 5N4Y as fact. If they'd really achieved that (with real, full nodes all going into full production), they wouldn't be where they are today. I'd happily never read about the supposed 5N4Y ever gain.

 

[MKWVentures]

I agree it probably won't be January.

But glass half full - there won't (likely?) be any retail TSMC N2 or Samsung 2nm products in Q1 2026 either..

Like meteor lake. they issue is not competition. For Intel, It is delivering to OEMs on the committed date

 

[Xebec]

I do hope the official history (if one ever exists) doesn't record Pat Gelsinger's claim of 5N4Y as fact. If they'd really achieved that (with real, full nodes all going into full production), they wouldn't be where they are today. I'd happily never read about the supposed 5N4Y ever gain.

This is kinda sticky because what is a real node these days? (danger of soap box incoming)

By TSMC's naming convention, N7 --> N3 is 5 distinct nodes (+4 nodes). Node 7, Node 6, Node 5, Node 4, Node 3. Yet in terms of density it's 2 nodes at most, and performance.. maybe as good as 1 node of Dennard scaling?

We'll have to see how much better transistor performance is for 18A vs. Intel early Intel 10nm (Icelake), when Pat took over.

 

[MKWVentures]

I do hope the official history (if one ever exists) doesn't record Pat Gelsinger's claim of 5N4Y as fact. If they'd really achieved that (with real, full nodes all going into full production), they wouldn't be where they are today. I'd happily never read about the supposed 5N4Y ever gain.

Intel 7 was a renamed 10nm node in development in 2021. Intel 4 volume was incredibly low. Meteor lake was a expensive blip . Intel 3 (essentially a 4+) is ramping and wont peak for a while. Intel 20 was cancelled. 18A isnt shipping to consumers yet (It is unclear when it will).

Intel 7 (Raptor Lake and Sapphire Rapids) are still the volume leaders (Can we please see IFS revenue by Node?)

 

[hist78]

Intel was able to do 5 nodes in 4 years because the lead fab has the mission to lead, seed and the seeds transfer to another fab. It used to be every other year a new node, before all the difficulty starting in the 2010 until present. The Ronler Acres Moore Park complex is a grand scale. There is probably a cheaper way to do R&D but that is the Intel way. It worked better when fab lite wasn't a viable option as it is now.

On the topic of taking fab lite off the table for Intel, so they focus on the old tick tock, now that Intel's capital structure is different, with Trump and SoftBank and Nvidia, perhaps they can sack the board members who still want to see fab lite happen? Make a statement that fab lite is dead?

In order to truly develop a viable process you need to ramp it to some level in order to find the most obvious flaws. That is what people are referring to when they talk about "yield learning". You want your R&D team involved in that process early on, so I believe that doing this makes sense. It also makes economic sense, because you have invested in installing a process line to produce the wafers during development anyway. Knowing that it will take several months to get the next fab up and running on the new process why wouldn't you use that capacity to start building inventory for your initial launch while you wait for the HVM factory to come on line.

While Oregon has multiple fabs, they are in essence all really one big fab. While kind of hard to see exactly the red ovals in this picture show links connecting the clean rooms.

Just my opinion, but I feel like Intel's costs are higher because historically they didn't care. The goal was technology leadership and if that cost more, so be it. We live in a different world now and if Intel is going to be successful, they need to learn some hard lessons, including cost consciousness and developing robust processes with larger process windows.

My question is: if Intel’s Oregon fab needs to play both R&D and production roles, would that affect its R&D mission? As we all know, semiconductor manufacturing is a time consuming process. If the R&D fab gets tied up in a 50-day production run, it could hurt R&D effectiveness. Intel’s competitors and customers are all moving forward at an amazing speed. Can Intel’s R&D fabs truly focus on R&D?

 

[hist78]

Intel 7 was a renamed 10nm node in development in 2021. Intel 4 volume was incredibly low. Meteor lake was a expensive blip . Intel 3 (essentially a 4+) is ramping and wont peak for a while. Intel 20 was cancelled. 18A isnt shipping to consumers yet (It is unclear when it will).

Intel 7 (Raptor Lake and Sapphire Rapids) are still the volume leaders (Can we please see IFS revenue by Node?)

Despite being labeled by Pat Gelsinger as technically and cost-wise uncompetitive, Intel’s 7 process node is facing capacity shortages due to strong demand for older Intel products based on it.

After Intel 7, what really happened to Intel 4, Intel 3, and Intel 20? Why are Intel’s OEM customers still willing to buy more products built on Intel 7? Something doesn’t seem right.

 

[Xebec]

Despite being labeled by Pat Gelsinger as technically and cost-wise uncompetitive, Intel’s 7 process node is facing capacity shortages due to strong demand for older Intel products based on it.

After Intel 7, what really happened to Intel 4, Intel 3, and Intel 20? Why are Intel’s OEM customers still willing to buy more products built on Intel 7? Something doesn’t seem right.

Short answer - Intel 7 was "full stack" - all server, workstation, desktop, and mobile use cases, and no "new node" has provided the volume or range of chips to fully replace everything Intel 7 provides.

Details:

Intel 4 produces mobile chips only (Meteor Lake), is of (very?) limited volume, and has a performance regression in some areas vs Intel 7 products, thanks to going from monolithic to chiplet/tile design (+ latency), and a clock speed regression (H/HX). So Intel 4 can't even fully replace Intel 7 mobile.

Intel 3 is mainly server chips right now, also (relatively) limited volume (Intel tends to manufacture server chips on multiple nodes anyway over time). Intel 3 can't fully replace Intel 7 for server products.

... Neither Intel 4 or 3 serve the desktop market in any significant way.

TSMC N3B - I assume is higher volume than Intel 4/3 for Intel products, but there are no server products, mobile and desktop only; desktop is "full stack", though no workstation class parts (i.e. Xeon W). mobile is more or less fully covered. OEMs were not impressed with Arrow Lake at launch.

Bonus 1 - Intel is replacing a lot of Raptor Lake products that failed - driving some more volume to Intel 7...

Bonus 2 - I've heard of rumors that Intel mandated OEMs buy a certain amount of Alder lake then Raptor lake (old Intel pressure style under Pat), so that may have obligated the OEMs to a certain amount of Intel 7 products even in 2025. Source is 'Moores Law is Dead', which Daniel Nenni has appeared on a few times.

Bonus 3 - Intel 7 has some 'niche' class products that no other node has, i.e. Alder Lake-N -- products based only on the e-cores from Alder Lake, such as Core i3-N305, Intel N97, etc.

 

[hist78]

Short answer - Intel 7 was "full stack" - all server, workstation, desktop, and mobile use cases, and no "new node" has provided the volume or range of chips to fully replace everything Intel 7 provides.

Details:

Intel 4 produces mobile chips only (Meteor Lake), is of (very?) limited volume, and has a performance regression in some areas vs Intel 7 products, thanks to going from monolithic to chiplet/tile design (+ latency), and a clock speed regression (H/HX). So Intel 4 can't even fully replace Intel 7 mobile.

Intel 3 is mainly server chips right now, also (relatively) limited volume (Intel tends to manufacture server chips on multiple nodes anyway over time). Intel 3 can't fully replace Intel 7 for server products.

... Neither Intel 4 or 3 serve the desktop market in any significant way.

TSMC N3B - I assume is higher volume than Intel 4/3 for Intel products, but there are no server products, mobile and desktop only; desktop is "full stack", though no workstation class parts (i.e. Xeon W). mobile is more or less fully covered. OEMs were not impressed with Arrow Lake at launch.

Bonus 1 - Intel is replacing a lot of Raptor Lake products that failed - driving some more volume to Intel 7...

Bonus 2 - I've heard of rumors that Intel mandated OEMs buy a certain amount of Alder lake then Raptor lake (old Intel pressure style under Pat), so that may have obligated the OEMs to a certain amount of Intel 7 products even in 2025. Source is 'Moores Law is Dead', which Daniel Nenni has appeared on a few times.

Bonus 3 - Intel 7 has some 'niche' class products that no other node has, i.e. Alder Lake-N -- products based only on the e-cores from Alder Lake, such as Core i3-N305, Intel N97, etc.

So, the “5 nodes in 4 years” (5N4Y) plan is actually one big node and four little ones over four years?

Since Intel introduced 5N4Y, I’ve always wondered how its OEM customers can keep up with such rapid cycles and manage their inventory accordingly. They could end up in serious overstock trouble if they’re not careful.

 

[Artificer60]

My question is: if Intel’s Oregon fab needs to play both R&D and production roles, would that affect its R&D mission? As we all know, semiconductor manufacturing is a time consuming process. If the R&D fab gets tied up in a 50-day production run, it could hurt R&D effectiveness. Intel’s competitors and customers are all moving forward at an amazing speed. Can Intel’s R&D fabs truly focus on R&D?

I'm not trying to be obtuse, but I can't even envision how you think the fab runs when you talk about getting "tied up in a 50-day production run". So let me give you a high level view of how the fab runs.

First you need to understand a few things about process development.

Even a new process uses 80%+ of the old equipment that was used in the last node. New tooling is only brought in when you have to deal with process geometries/materials that the old process equipment can't manage. As much as possible you work on developing new recipes on the existing equipment along with the new masks that you need with the smaller features.

You don't start very many development wafers. You don't need very many wafers to find the bigger process cliffs. That means you don't need very many tools for early development. Tools dedicated to development often sit idle for extended periods waiting for the next batch of development wafers. As you get closer to your process goals you need more volume to find the defects that occur less frequently and you need more tooling. But you want to start the smallest number of development wafers you can and still get the job done, because every development wafer you make is money down the drain. You will never sell those wafers so waste as little cash on that as you can.

The consequences of the previous statements means that when you start developing a new process, except for the new tools, you have a lot more capacity than you need. That capacity is completely capable of running the previous process that was developed on it. So you can either run the old process, or let the tools sit idle. And an idle fab is probably the most expensive piece of property on the planet. So you want to keep all your tools busy and that means running the old process to fill this capacity.

When you are planning how to run the fab you know your capacity. You allocate a certain amount of that capacity for development wafers and the rest of it for products on your other wafers. All those other wafers can be comprised of multiple different products and different quantities for each of those products. All your wafers have a target completion date (though your development wafers may not run through the full process, because you don't need the whole process to evaluate many process issue and we aren't looking to waste money). Given that you know how long it takes a wafer to move through your factory and you know when you have to deliver it, you can figure out when you need to start the wafer. So you have allocated some amount of capacity to the technology wafers and the rest of the capacity is filled with the other wafers. It is the need to reserve capacity for the TD wafers that causes the Oregon fabs from generating the same level of output that the other factories in Intel's factory network can run.

You can resolve the focus problem with two groups of engineers. One group that sustains the tools and makes small process tweaks to improve the older processes and another group that focuses on resolving the issues that arise during development.

 

[hist78]

I'm not trying to be obtuse, but I can't even envision how you think the fab runs when you talk about getting "tied up in a 50-day production run". So let me give you a high level view of how the fab runs.

First you need to understand a few things about process development.

Even a new process uses 80%+ of the old equipment that was used in the last node. New tooling is only brought in when you have to deal with process geometries/materials that the old process equipment can't manage. As much as possible you work on developing new recipes on the existing equipment along with the new masks that you need with the smaller features.

You don't start very many development wafers. You don't need very many wafers to find the bigger process cliffs. That means you don't need very many tools for early development. Tools dedicated to development often sit idle for extended periods waiting for the next batch of development wafers. As you get closer to your process goals you need more volume to find the defects that occur less frequently and you need more tooling. But you want to start the smallest number of development wafers you can and still get the job done, because every development wafer you make is money down the drain. You will never sell those wafers so waste as little cash on that as you can.

The consequences of the previous statements means that when you start developing a new process, except for the new tools, you have a lot more capacity than you need. That capacity is completely capable of running the previous process that was developed on it. So you can either run the old process, or let the tools sit idle. And an idle fab is probably the most expensive piece of property on the planet. So you want to keep all your tools busy and that means running the old process to fill this capacity.

When you are planning how to run the fab you know your capacity. You allocate a certain amount of that capacity for development wafers and the rest of it for products on your other wafers. All those other wafers can be comprised of multiple different products and different quantities for each of those products. All your wafers have a target completion date (though your development wafers may not run through the full process, because you don't need the whole process to evaluate many process issue and we aren't looking to waste money). Given that you know how long it takes a wafer to move through your factory and you know when you have to deliver it, you can figure out when you need to start the wafer. So you have allocated some amount of capacity to the technology wafers and the rest of the capacity is filled with the other wafers. It is the need to reserve capacity for the TD wafers that causes the Oregon fabs from generating the same level of output that the other factories in Intel's factory network can run.

You can resolve the focus problem with two groups of engineers. One group that sustains the tools and makes small process tweaks to improve the older processes and another group that focuses on resolving the issues that arise during development.

From your observation, does Intel not have a need for a fully dedicated R&D fab? Or does Intel simply not have enough R&D projects or funding to fully utilize its Oregon R&D fabs?

I'm not sure TSMC ever counts its R&D fab as a production site. Given its rapid deployment schedule, I don’t think it can afford to do that.

 

[Artificer60]

According to Google, you are making a flawed assumption. When asked if TSMC dedicates a fab to development, this is the response.

"No, TSMC does not dedicate an entire fab exclusively to development; instead, it uses a hybrid approach. While the majority of a factory is used for volume production, a portion of the fab is allocated for research and development (R&D) purposes for the next process node. When the new process is ready, it is transferred to other fabs for high-volume manufacturing"

 

[dkr1986]

Despite being labeled by Pat Gelsinger as technically and cost-wise uncompetitive, Intel’s 7 process node is facing capacity shortages due to strong demand for older Intel products based on it.

But the title of the chart says Two Fabs started 18A production.

The answer to these could be Fab 42 (EUV capable) in Arizona which was Intel 7 is being upgraded to Intel 18A. So, Intel 7 will be capacity constrained for foreseeable future. Lot of write down on Q3'24 earning was related Intel 7 equipment that probably did not have reuse. So Intel could be ramping up Fab 42 for Intel 18A production too (or they could be just referring to Oregon R&D fab that can output some limited wspm).

Fab 62 will be more likely "shell'ed" until external customer volume on 18A ramps up.

IIRC, Intel was running Intel 7 on both Fab 42 at Arizona & Fab 28 at Israel during 2022 to 2024. Going forward It looks like Fab 28 at Israel will be the only Intel 7 fab. So, when Intel Products transition fully to 18A, they need 2 fabs worth of capacity at least. If people want to keep claiming Intel 18A yield is mediocre, then they need more capacity as they need to run more wafers to meet same product output. It will not be good financially but that is what low yield on 18A implies (I don't believe that)

Something to keep in mind is 18A Fabs at Arizona campus can be augmented by older fabs there that can run DUV steps etc. & are connected by FOUPS network.

https://download.intel.com/newsroom/2024/corporate/Intel-Arizona-The-Silicon-desert.pdf

After Intel 7, what really happened to Intel 4, Intel 3, and Intel 20? Why are Intel’s OEM customers still willing to buy more products built on Intel 7? Something doesn’t seem right.

It is in the best interest of Intel as a business to keep Intel 4 based Meteor Lake volume down due to major outsourcing of tiles to TSMC. Only the compute tile which is tiny is on Intel 4, rest of the tiles are on TSMC nodes. The same story will Arrow Lake which is fully TSMC. So, I presume Intel is intentionally keeping volumes down on these. Exception to these is Lunar Lake which is a huge success per Intel that it went from a niche product to increased volume as much as 3x originally planned.

Intel 3 imo is doing well, it is used in GNR & SRF server products (& ARL-U mobile laptop chips) that are ramping up nicely. It will take another 4 Qtrs-6Qtrs for those to reach peak volume (right now Sapphire Rapids & Emerald Rapids, both on Intel 7 is the major volume driver for Intel DC server CPUs). Intel 3 will also be used to base tiles for upcoming Clearwater Forrest (& I assume Diamond rapids too in end of 2026). Part of iGPU tiles on Panther Lake is also on Intel 3.

Intel 20A productization was cancelled. so, no products.

imo, Intel 18A is the most important node for Intel. They need to accelerate moving all products to Intel 18A asap. That should help on both Gross Margin (after the node ramps - there will initial ramp cost - all nodes do) & ASP perspective. Only time will tell us how that goes.

 

[Xebec]

So, the “5 nodes in 4 years” (5N4Y) plan is actually one big node and four little ones over four years?

Since Intel introduced 5N4Y, I’ve always wondered how its OEM customers can keep up with such rapid cycles and manage their inventory accordingly. They could end up in serious overstock trouble if they’re not careful.

Hah - one big + 4 little is a fair intepretation, if we're talking volume . 1 of those (20A) is even a phantom node... (Though I hope a year from now we'll be looking at 18A as a solid process that put Intel ahead of Samsung, and back in real competition +/- with TSMC. This is a large technical improvement from 2016-2020).

re: Inventory - it IS a mess, and I think OEMs are currently hurting from "Intel 7" overstock thanks to the failing Raptor Lake CPUs that both lowered demand and required returns/replacements.

That said, I've often heard that the OEMs and retailers are the ones often asking for annual updates / refreshes so there is 'something new and shiny' to sell, but it looks the wind is starting to change to every 2 years for commodity parts (GPUs, AMD CPUs), and even longer for some platforms.

 

[tomatoma]

The answer to these could be Fab 42 (EUV capable) in Arizona which was Intel 7 is being upgraded to Intel 18A. So, Intel 7 will be capacity constrained for foreseeable future. Lot of write down on Q3'24 earning was related Intel 7 equipment that probably did not have reuse. So Intel could be ramping up Fab 42 for Intel 18A production too (or they could be just referring to Oregon R&D fab that can output some limited wspm).

Fab 62 will be more likely "shell'ed" until external customer volume on 18A ramps up.

IIRC, Intel was running Intel 7 on both Fab 42 at Arizona & Fab 28 at Israel during 2022 to 2024. Going forward It looks like Fab 28 at Israel will be the only Intel 7 fab. So, when Intel Products transition fully to 18A, they need 2 fabs worth of capacity at least. If people want to keep claiming Intel 18A yield is mediocre, then they need more capacity as they need to run more wafers to meet same product output. It will not be good financially but that is what low yield on 18A implies (I don't believe that)

Something to keep in mind is 18A Fabs at Arizona campus can be augmented by older fabs there that can run DUV steps etc. & are connected by FOUPS network.

https://download.intel.com/newsroom/2024/corporate/Intel-Arizona-The-Silicon-desert.pdf

View attachment 3750


It is in the best interest of Intel as a business to keep Intel 4 based Meteor Lake volume down due to major outsourcing of tiles to TSMC. Only the compute tile which is tiny is on Intel 4, rest of the tiles are on TSMC nodes. The same story will Arrow Lake which is fully TSMC. So, I presume Intel is intentionally keeping volumes down on these. Exception to these is Lunar Lake which is a huge success per Intel that it went from a niche product to increased volume as much as 3x originally planned.

Intel 3 imo is doing well, it is used in GNR & SRF server products (& ARL-U mobile laptop chips) that are ramping up nicely. It will take another 4 Qtrs-6Qtrs for those to reach peak volume (right now Sapphire Rapids & Emerald Rapids, both on Intel 7 is the major volume driver for Intel DC server CPUs). Intel 3 will also be used to base tiles for upcoming Clearwater Forrest (& I assume Diamond rapids too in end of 2026). Part of iGPU tiles on Panther Lake is also on Intel 3.

Intel 20A productization was cancelled. so, no products.

imo, Intel 18A is the most important node for Intel. They need to accelerate moving all products to Intel 18A asap. That should help on both Gross Margin (after the node ramps - there will initial ramp cost - all nodes do) & ASP perspective. Only time will tell us how that goes.

I think we'll be seeing Intel 3 in many different situations in the future.

 

[tomatoma]

According to Google, you are making a flawed assumption. When asked if TSMC dedicates a fab to development, this is the response.

"No, TSMC does not dedicate an entire fab exclusively to development; instead, it uses a hybrid approach. While the majority of a factory is used for volume production, a portion of the fab is allocated for research and development (R&D) purposes for the next process node. When the new process is ready, it is transferred to other fabs for high-volume manufacturing"

For now, it seems that fab12B is clearly being used for research and development.

 

[tomatoma]

I'm not trying to be obtuse, but I can't even envision how you think the fab runs when you talk about getting "tied up in a 50-day production run". So let me give you a high level view of how the fab runs.

First you need to understand a few things about process development.

Even a new process uses 80%+ of the old equipment that was used in the last node. New tooling is only brought in when you have to deal with process geometries/materials that the old process equipment can't manage. As much as possible you work on developing new recipes on the existing equipment along with the new masks that you need with the smaller features.

You don't start very many development wafers. You don't need very many wafers to find the bigger process cliffs. That means you don't need very many tools for early development. Tools dedicated to development often sit idle for extended periods waiting for the next batch of development wafers. As you get closer to your process goals you need more volume to find the defects that occur less frequently and you need more tooling. But you want to start the smallest number of development wafers you can and still get the job done, because every development wafer you make is money down the drain. You will never sell those wafers so waste as little cash on that as you can.

The consequences of the previous statements means that when you start developing a new process, except for the new tools, you have a lot more capacity than you need. That capacity is completely capable of running the previous process that was developed on it. So you can either run the old process, or let the tools sit idle. And an idle fab is probably the most expensive piece of property on the planet. So you want to keep all your tools busy and that means running the old process to fill this capacity.

When you are planning how to run the fab you know your capacity. You allocate a certain amount of that capacity for development wafers and the rest of it for products on your other wafers. All those other wafers can be comprised of multiple different products and different quantities for each of those products. All your wafers have a target completion date (though your development wafers may not run through the full process, because you don't need the whole process to evaluate many process issue and we aren't looking to waste money). Given that you know how long it takes a wafer to move through your factory and you know when you have to deliver it, you can figure out when you need to start the wafer. So you have allocated some amount of capacity to the technology wafers and the rest of the capacity is filled with the other wafers. It is the need to reserve capacity for the TD wafers that causes the Oregon fabs from generating the same level of output that the other factories in Intel's factory network can run.

You can resolve the focus problem with two groups of engineers. One group that sustains the tools and makes small process tweaks to improve the older processes and another group that focuses on resolving the issues that arise during development.

Generally, the research and development team is divided into several large teams.
After-sales support for current processes, development of next-generation processes, and development of processes of the next generation...

 

[MKWVentures]

Despite being labeled by Pat Gelsinger as technically and cost-wise uncompetitive, Intel’s 7 process node is facing capacity shortages due to strong demand for older Intel products based on it.

After Intel 7, what really happened to Intel 4, Intel 3, and Intel 20? Why are Intel’s OEM customers still willing to buy more products built on Intel 7? Something doesn’t seem right.

Its all about price and value and to some degree cost. Raptor lake is a good product at a good price. and OEMs see value in it. The margins are higher than Intel 4.... and we will soon see whether they are still better than 18A. I would love to see the OEM BOM cost for a raptor lake, meteor lake, lunar lake, Panther lake Notebook. Personally, I love my $450 notebooks. My alder lake and raptor lake notebooks run everything I need including AI apps ( obviously I am not a power user).

Datacenter is ALWAYS way slower than people realize to convert. I think granite rapids is still over a year from being at peak volume

Hopefully Intel add some details on volume by node this month

 

[siliconbruh999]

Its all about price and value and to some degree cost. Raptor lake is a good product at a good price. and OEMs see value in it. The margins are higher than Intel 4.... and we will soon see whether they are still better than 18A. I would love to see the OEM BOM cost for a raptor lake, meteor lake, lunar lake, Panther lake Notebook. Personally, I love my $450 notebooks. My alder lake and raptor lake notebooks run everything I need including AI apps ( obviously I am not a power user).

Intel 4 is small part of MTL lake and the rest is TSMC and OEMs still like RPL despite it's flaws. In credit to RPL it clocks like 6Ghz and 6.2 Ghz on binned and it gives Intel N3B ARL and N4P Zen 5 run for it's money though using more power.