LNL is hilariously more efficient than ARL in the intended scenarios for LNL:
- On package memory
- IMC on compute die
- Fewer D2D power losses
- Higher percentage of total Si area using the most advanced process
- Better NPU and GPU IPs are more efficient
- Smaller GPU (more power efficient)
- GPU, NPU, LP e core cluster on same die with CPU
- SOC cache
- Better power delivery and regulators
- Finally phasing out Intel's outdated race to idle power management scheme for a more mobile like rapid switching between various power states to minimize power consumption across the workload rather than after the work was done
As for Pantherlake only being similar to LNL on battery life. There are choices that Intel communicated, or are wildly known or believed to be true that would blur some of the improvements you would expect from say a 2nm LNL:
- Intel every quarter talks about how no on package memory saves a ton of cash and lifts margins
- PNL has more dies increasing power consumption
- It is unclear if PNL has SOC cache still
- Core IP is a die shrink so it isn't meaningfully better if you take away the extra 18A process entitlment
- Rumored to have a bigger GPU than LNL
- This is first year of production 18A vs second year of production N3 so N3 variation should be lower (reducing some of the power advantages 18A intrinsically has)
- The whole chip isn't 18A so the 18A benefits only apply to the 18A components
- Since this wasn't intended to be an ultra premium MacBook air compete chip it likely won't have those fancy PMICs
- PNL has a higher performance target than LNL so not all of the power savings are being passed towards lowering power consumption
- It is unknown if PNL is race to idle or not, but if it is whoever made that call needs to be fired ASAP. So I would tentatively say it should have that
- While 18A is pretty close to N2, N3 is also pretty close to N2. So it isn't like N3 to 18A was ever going to be some giant leap (like say going from T40nm to T28nm)
To Intel products it doesn't matter if utilization is 1%. They pay a fixed price and MJ says it is cheaper and higher margin than current products. Especially LNL which is already being sold into cheaper laptops than expected because ARL U (and no I don't mean MTL on Intel 3) was that bad and nobody wanted it. Pictures also seems like die area has walked away from the near 300mm2 of ARL and more than 250mm2 MTL. If the biggest PNL die for the biggest sku is around 140, then whole chips is definitely smaller than MTL.
As for foundry: clearly ramp it isn't unbearably painful since this is the peak of foundry losses (all the cost plus no revenue until EOY), and Intel has continued to stick with the story of 27 break even and rapid operating loss reduction in 2026.
I mean yeah. How could it not? It is a fact Intel 7 is sold near cost, and Intel says the 18A ASP is 3x while wafer cost is not up meaningfully. 2+2=4 and 18A margin is greatly improved. If 18A margins weren't at least ok, break even would have never been possible in 2027. Let's be bearish and say single digit Intel 7 margin means 1%. Cost = 1 and price = 1.01. Even if we say Intel is lying and 18A cost isn't modestly higher, but is really 2x i7. Okay P=3.03 and C=2. GM is 34% (lower than what TSMC makes on these new nodes but similar to what UMC or GF do in fully deprecated fabs). Even if we say i7 GM is -9% (P=1 and C=1.09), 18A margin comes out as 27% with all other assumptions being the same (3x price and 2x cost).
edit: fixed the math for the 1% margin case. I had a brain fart and used the wrong number and got a 39% margin rather than a 34%.
As a side note the statement of 18A cost not being much higher than i7 and being similar to N2 means that Intel 7 cost is somewhere around N3 cost of production. Considering a lot of those N3 layers are more expensive EUV layers, that would indicate that i7 probably has wildly more mask layers than even N2 if cost is similar to N3. Which is insane but not exactly shocking when you look at a teardown. It certainly explains why for all of the Intel 7 volume there is no profit after you take out R&D, ecosystem building, and building more than 1 EUV capable fab outside of Oregon. One can certainly see why Intel is very jazzed to be moving past i7.
.