Intel will ship the world's first "2nm class" GPU?

[Brady]

Intel will introduce Panther Lake on the 18A "2nm class" process node featuring the next-gen Celestial (Xe3) GPU with a purported 12 Xe Cores (current top Battlemage, the discrete B580, has 20 Xe Cores). With the latest quarterly earnings update, Intel confirmed PTL for 25H2 "release" though I'm sure the ramp will be long and throughout 2026.

Does this mean they will beat both AMD and Nvidia to 2nm? Is the GPU tile on 18A or TSMC still?

Per More Than Moore:

Complicating matters, Intel’s product groups have opted to use rival TSMC’s 3nm process nodes for both of Intel’s current-gen (200 series) Core processors, Lunar Lake and Arrow Lake, so Intel’s fabs are not enjoying as much business from themselves as they’d normally see. Co-CEO Michelle Johnston Holthaus says that for every generation of product, they need the right product on the right node for the right time - and so Foundry has to win the Products business. That’s going to be true for Panther Lake, as 70% of that silicon will be in house - as part of the Q&A Holthaus stated that Nova Lake, the generation beyond, will also be a mix of foundry offerings. Nonetheless, back to the numbers here, Q4 foundry revenue was down 13% year over year, and Lunar/Arrow are still ramping.

Apple appears to not be making the jump to N2 for the upcoming M5 chip family.

Regardless of the relatively small size of the PTL GPU tile, it seems like a watershed moment that Intel could be the first 2nm GPU to ship to customers, perhaps even holding that title deep into 2026. With Intel seeming to have really hit their stride with Battlemage, this could be a big moment for their beleaguered parallel computing ambitions. Are there any rumors yet of discrete Xe3? Or if Jaguar Shores will use that or Xe4?

This is also of course expected to be the make or break moment for Intel Foundry; simply put, 18A needs to excel and PTL needs to shine as a product.

Fun times ahead.

 

[siliconbruh999]

Intel will introduce Panther Lake on the 18A "2nm class" process node featuring the next-gen Celestial (Xe3) GPU with a purported 12 Xe Cores (current top Battlemage, the discrete B580, has 20 Xe Cores). With the latest quarterly earnings update, Intel confirmed PTL for 25H2 "release" though I'm sure the ramp will be long and throughout 2026.

Does this mean they will beat both AMD and Nvidia to 2nm? Is the GPU tile on 18A or TSMC still?

Per More Than Moore:



Apple appears to not be making the jump to N2 for the upcoming M5 chip family.

Regardless of the relatively small size of the PTL GPU tile, it seems like a watershed moment that Intel could be the first 2nm GPU to ship to customers, perhaps even holding that title deep into 2026. With Intel seeming to have really hit their stride with Battlemage, this could be a big moment for their beleaguered parallel computing ambitions. Are there any rumors yet of discrete Xe3? Or if Jaguar Shores will use that or Xe4?

This is also of course expected to be the make or break moment for Intel Foundry; simply put, 18A needs to excel and PTL needs to shine as a product.

Fun times ahead.

The GPU For Panther Lake is dual sourced with N3E and Intel 3 as for First 2nm GPU I think it is something we will know in 2026

 

[Brady]

Betteridge's Law of Headlines strikes again.

 

[Daniel Nenni]

Betteridge's Law of Headlines strikes again.

I had to look it up

Betteridge's law of headlines - Wikipedia

en.wikipedia.org

 

[MKWVentures]

Nvidia and AMD have never had their product on the leading node. That is not the goal. Success is the goal.

Client GPU is not the same as DC GPU. blurring the two is not recommended. "AI PC" is not DC AI

 

[XYang2023]

Nvidia and AMD have never had their product on the leading node. That is not the goal. Success is the goal.

Client GPU is not the same as DC GPU. blurring the two is not recommended. "AI PC" is not DC AI

I disagree. The B580 can out-compete some exiting data center GPUs. I think the underlying IPs were also meant for data centers when they were designed.

 

[Xebec]

Nvidia and AMD have never had their product on the leading node. That is not the goal. Success is the goal.

Client GPU is not the same as DC GPU. blurring the two is not recommended. "AI PC" is not DC AI

I thought AMD always had CPUs on leading edge nodes for decades up until their old fabs/GF missed at 7nm?

AMD also had Radeon 7900 series on TSMC 28nm in 2011, when it was their most advanced node. (The GPUs have been a bit less aggressive since then when Lisa Su favored Zen CPU funding over GPU).

There have also been some rumors that Nvidia Blackwell was originally intended for N3. It's also very large and pretty power hungry on N4.

 

[MKWVentures]

I thought AMD always had CPUs on leading edge nodes for decades up until their old fabs/GF missed at 7nm?

AMD also had Radeon 7900 series on TSMC 28nm in 2011, when it was their most advanced node. (The GPUs have been a bit less aggressive since then when Lisa Su favored Zen CPU funding over GPU).

There have also been some rumors that Nvidia Blackwell was originally intended for N3. It's also very large and pretty power hungry on N4.

The whole world changed when TSMC took the lead. So the history is like another planet IMO. AMD and Nvidia are always one gen back (but still super competitive). If you look how its managed by TSMC it i real brilliance on capital usage. TSMC also uses pricing to get what they want on distribution across nodes.. I just want to keep reminding people that the worlds most hyped processor (Blackwell) is on N4 and it will not peak until at least the end of the year..... N2 will be running at that time.

 

[Xebec]

The whole world changed when TSMC took the lead. So the history is like another planet IMO. AMD and Nvidia are always one gen back (but still super competitive). If you look how its managed by TSMC it i real brilliance on capital usage. TSMC also uses pricing to get what they want on distribution across nodes.. I just want to keep reminding people that the worlds most hyped processor (Blackwell) is on N4 and it will not peak until at least the end of the year..... N2 will be running at that time.

Fair - just my gut is just telling me N3 capacity is a bit constrained* if we don't see any AMD or Nvidia products on it after (more than) 3 years of production.. (Exception : Zen 5 compact cores). Based on current GPU and CPU cycles of 20-24 months releases, that means no major N3 for AMD or Nvidia until Year 5 of production.

*Or the price/performance isn't compelling enough vs N4/N5.

 

[Daniel Nenni]

The whole world changed when TSMC took the lead. So the history is like another planet IMO. AMD and Nvidia are always one gen back (but still super competitive). If you look how its managed by TSMC it i real brilliance on capital usage. TSMC also uses pricing to get what they want on distribution across nodes.. I just want to keep reminding people that the worlds most hyped processor (Blackwell) is on N4 and it will not peak until at least the end of the year..... N2 will be running at that time.

Well said. The best is yet to come for TSMC N3. It will be a very long node with the most total tape-outs in the history of TSMC, my opinion.

Fair - just my gut is just telling me N3 capacity is a bit constrained* if we don't see any AMD or Nvidia products on it after (more than) 3 years of production.. (Exception : Zen 5 compact cores). Based on current GPU and CPU cycles, that means no major N3 for AMD or Nvidia until Year 5 of production.
*Or the price/performance isn't compelling enough vs N4/N5.

TSMC built N3 capacity knowing what demand would be. In fact, given that Intel did not fully use what they originally planned, N3 capacity should not be a problem. Pricing is definitely a factor for AMD and Nvidia for sure but I think their node cadence is normal. These are very complex chips and are mostly new designs whereas Apple's SoCs are more iterative based on previous designs.

Back in the day we used to do shrinks of existing designs. Not so much anymore.

 

[kevin01]

The GPU For Panther Lake is dual sourced with N3E and Intel 3 as for First 2nm GPU I think it is something we will know in 2026

Fair - just my gut is just telling me N3 capacity is a bit constrained* if we don't see any AMD or Nvidia products on it after (more than) 3 years of production.. (Exception : Zen 5 compact cores). Based on current GPU and CPU cycles of 20-24 months releases, that means no major N3 for AMD or Nvidia until Year 5 of production.

*Or the price/performance isn't compelling enough vs N4/N5.

Nvda gpu dies are very large, so from a yield point of view, very mature N4P is probably more suitable than N3.

 

[siliconbruh999]

Well said. The best is yet to come for TSMC N3. It will be a very long node with the most total tape-outs in the history of TSMC, my opinion.



TSMC built N3 capacity knowing what demand would be. In fact, given that Intel did not fully use what they originally planned, N3 capacity should not be a problem. Pricing is definitely a factor for AMD and Nvidia for sure but I think their node cadence is normal. These are very complex chips and are mostly new designs whereas Apple's SoCs are more iterative based on previous designs.

Back in the day we used to do shrinks of existing designs. Not so much anymore.

So Intel booked a lot more TSMC Capacity for N3?
Jesus Intel I don't know what they want to do now since they are migrating many things to Intel nodes from GPU/SOC/CPU.

 

[Xebec]

TSMC built N3 capacity knowing what demand would be. In fact, given that Intel did not fully use what they originally planned, N3 capacity should not be a problem. Pricing is definitely a factor for AMD and Nvidia for sure but I think their node cadence is normal. These are very complex chips and are mostly new designs whereas Apple's SoCs are more iterative based on previous designs.

Back in the day we used to do shrinks of existing designs. Not so much anymore.

The shrink days definitely seemed more straight forward . It was awesome seeing "free" frequency and/or power.

Who is shipping products based on the N3 capacity these days besides Apple and Intel then?

I think this is the timeline:

2022 - N3 enters mass production
2023 - Apple ships first N3 products (iPhone Pro A-series and M3 chip), but also releases new N4 products (iPhone 'regular')
2024 - Apple releases more N3 products, AMD has limited N3 release via Zen 5C, Intel launches mainstream N3 products in 2H;.
Q4 2024 - Apple launches "big die" N3 products: M3 Pro, M3 Max.

2025 - No new N3 from AMD or Nvidia; Intel continues ramp on N3 (but some A18 in Q4), Apple releases M5 chip on TSMC N3, new iPhone TBD but probably N3 not N2. Maybe Qualcomm has N3 products this year?

(2025 - N2 Enters Mass Production)

2026 - Tesla AI5 hardware rumored Q4 2025/Q1 2026 production on N3, AMD Zen 6 on N3 summer, Radeon and Nvidia GPUs on N3 in Q4

(2026 - First N2 Apple products)

 

[Xebec]

Nvda gpu dies are very large, so from a yield point of view, very mature N4P is probably more suitable than N3.

Fair. FWIW Apple does have a >500mm2 die (maybe 600mm2) working on N3 since Oct 2024 -- the Apple M3 Ultra.

 

[nghanayem]

So Intel booked a lot more TSMC Capacity for N3?
Jesus Intel I don't know what they want to do now since they are migrating many things to Intel nodes from GPU/SOC/CPU.

Yeah, why wouldn't they have excess N3 capacity. Back in 2020 when the wafer agreements were being signed, intel was expecting pandemic era PC unit shipment growth to continue. Throw in intel last minute shifting substantial demand back internal, and Intel product's demand projections only being around 3/5 of what Intel thought it should be in that 2021-2022 timing. How could Intel possibly consume all of the wafers that they would have originally planned for in 2020. Looking back, I suspect that Pat's various flights out to TSMC back in 2022 were negotiating reductions to Intel's N6, N5P, N4P, N3B, etc. wafer agreements.

Well said. The best is yet to come for TSMC N3. It will be a very long node with the most total tape-outs in the history of TSMC, my opinion.

One can say that about almost every new TSMC node across their history . 28nm was the biggest node ever, then 16FF became a bigger node, then it was N7, then N5, and now N3. Looking at the number of fabs being built, it looks like TSMC plans to continue their trend and have N2 be the next biggest node in TSMC history.

The shrink days definitely seemed more straight forward . It was awesome seeing "free" frequency and/or power.

Who is shipping products based on the N3 capacity these days besides Apple and Intel then?

I think this is the timeline:

2022 - N3 enters mass production
2023 - Apple ships first N3 products (iPhone Pro A-series and M3 chip), but also releases new N4 products (iPhone 'regular')
2024 - Apple releases more N3 products, AMD has limited N3 release via Zen 5C, Intel launches mainstream N3 products in 2H;.
Q4 2024 - Apple launches "big die" N3 products: M3 Pro, M3 Max.

2025 - No new N3 from AMD or Nvidia; Intel continues ramp on N3 (but some A18 in Q4), Apple releases M5 chip on TSMC N3, new iPhone TBD but probably N3 not N2. Maybe Qualcomm has N3 products this year?

(2025 - N2 Enters Mass Production)

2026 - Tesla AI5 hardware rumored Q4 2025/Q1 2026 production on N3, AMD Zen 6 on N3 summer, Radeon and Nvidia GPUs on N3 in Q4

(2026 - First N2 Apple products)

QCOM launched X-elite based smartphone SOCs at the end of 2024 for the 2025 smartphone refreshes. MediaTek announced an N3E chip last year, but to my knowledge, nothing has launched with it yet. My guess is you will start seeing them in Motorola's second from the top spec phones this summer and trickling down to various manufacture's lower end devices over the coming couple of years. I wouldn't be too rough on TSMC though. Even though 2023 is when N3 products launched, effectively the action didn't really start until 2024 since original N3 was kind of DOA and folks needed the extra time to move their IPs from N3 to the N3E ecosystem while TSMC got that new process to better than N3 defect densities. The whole odyssey caused something of a false start if you will. N2 should go MUCH smoother since TSMC didn't need to redefine their node and throw out their whole FEOL and MEOL process to start from scratch.

Nvda gpu dies are very large, so from a yield point of view, very mature N4P is probably more suitable than N3.

CC Wei did say back in like 2022 or 2023 that process complexity was getting to the point that the time it takes to hit various defect density levels would increase. As shown by the time from 10FF to N5 being 3 years, while the time between N5 and N2 is 6 years. Lengthening cycle times also make this matter even worse, since it takes you longer to get the results of your various experiments. His statement was in reference to N2, but the same principles apply to N3 family (of course more heavily to N3 than N3E as N3E isn't that much more complex than N5). Add more mask layers, and your DD for each layer needs to be lower to hit a given completed wafer DD.

Fair - just my gut is just telling me N3 capacity is a bit constrained* if we don't see any AMD or Nvidia products on it after (more than) 3 years of production.. (Exception : Zen 5 compact cores). Based on current GPU and CPU cycles of 20-24 months releases, that means no major N3 for AMD or Nvidia until Year 5 of production.

*Or the price/performance isn't compelling enough vs N4/N5.

For reference, NVIDIA didn't launch N5/4 parts until 2022 (2 years after first N5 iphones) and their 2020 datacenter parts were the first N7 family chips they did (2 years after first N7 iphones). If NVIDIA doesn't have any N3 parts until 2026 that is just them doing business as usual (2 years after N3E). Bigger chips take longer to design and are harder to yield. Also given AMD can't even pull out unquestioned efficiency leads even with a process lead, NVIDIA doesn't exactly have much of an impetus to move to a newer node until it will make their products cheaper (both from slowly declining pricing and from improving yields).

 

[Daniel Nenni]

One can say that about almost every new TSMC node across their history . 28nm was the biggest node ever, then 16FF became a bigger node, then it was N7, then N5, and now N3. Looking at the number of fabs being built, it looks like TSMC plans to continue their trend and have N2 be the next biggest node in TSMC history.

Not true at all. TSMC N3 was unopposed and had a much larger design win market share than previous nodes. Intel 3 and Samsung 3nm had zero foundry market share. I have never seen this phenomenon in my career and hopefully it will not happen again. TSMC N2 will compete against Intel 18A and Samsung 2nm so it will not be the same.

 

[siliconbruh999]

Not true at all. TSMC N3 was unopposed and had a much larger design win market share than previous nodes. Intel 3 and Samsung 3nm had zero foundry market share. I have never seen this phenomenon in my career and hopefully it will not happen again. TSMC N2 will compete against Intel 18A and Samsung 2nm so it will not be the same.

Well when the competition is your client you are going to win now the competition has it's own thing working so 1 big customer is gone I just hope 18A gets clients and Samsung as well

 

[Brady]

Well, now we have a rumor of “Xe3P” architecture being manufactured in-house: https://videocardz.com/newz/intels-...to-feature-xe3p-architecture-may-not-use-tsmc

This is probably a late 2026 thing at the earliest, but it would be a “2nm class” DGPU if this rumor is true.