A report from Morgan Stanley makes some really bold claims. According to the analyst, by 2024, roughly 20-25% of Intel CPUs will be fabbed by TSMC (primarily) on its 3nm-class process nodes. This is highly doubtful as Team Blue plans on manufacturing most of its CPU tiles by itself. Meteor and Arrow Lake which be the first consumer processors to leverage a chiplet (MCM) design, will be fabbed using the Intel 7 and 20A (2nm) process nodes, with TSMC’s N3 used for the iGPU chiplet.
Considering that Meteor Lake and Arrow Lake will have overlapping product cycles, TSMC will certainly have a part to play in Intel’s microprocessors from 2023 to 2025, at the very least. Going by Intel’s own block diagrams (which may be just placeholders), the compute dies will be fabbed on its internal nodes, leaving TSMC with either the SoC or the GPU die.
Outsourcing the I/O die to an expensive process node like N3 just isn’t practical which means TSMC will most likely produce Intel’s iGPU dies from 2023 onwards. These will be fabbed on an N3-class process node which is progressing quite well. According to the latest report, N3e production yields continue to improve, with the volume production reportedly pulled ahead of schedule. It’s now slated to start in the second quarter of 2023, from the earlier set third quarter of 2023. Morgan Stanley claims that the logic density of N3e is just 8% less than N3 (by cutting four UV layers), but a considerable 60% denser than N5.
Via: Dylan Patel