If Lip Bu Tan can't sell you his LLM accelerator, he's more than willing to build yours
theregister.com, Apr. 30, 2025 –
Intel has revealed a pair of variants of its long-awaited 18A process node to make it better suited for, one, manufacturing mass-market processors and, two, complex multi-die semiconductors for – of course – AI.
First teased in mid-2021, the 2nm-ish 18A is set to finally enter volume production later this year with the launch of Intel’s Panther Lake client processor family. However, the node was only really intended for specialist high-performance use cases — think high-end CPUs and GPUs — with its forthcoming 14A process node set to be the chipmaker’s first truly mass-market node.
Unfortunately, while 14A promises to deliver a 15-20 percent performance per watt uplift, it’s still a few years away from volume production.
Faced with growing demand, particularly for American-made silicon, Intel has decided it needs something to satiate the masses a little sooner and has tweaked its original 18A formula with two new revisions: One designed to address a broad range of applications, and another tuned for, you guessed it, multi-die AI accelerators.
The first of these is called 18A-P, and promises another 8 percent performance per watt improvement over the base 18A node. More importantly, for customers already evaluating designs on 18A, Intel Chief Global Operations Officer Naga Chandrasekaran claimed transitioning to the newer node should be seamless.
The other variant is called 18A-PT and is optimized for designs requiring through silicon vias (TSVs), which allow communication and power delivery through silicon and means chips can be assembled by stacking them on top of one another.
At the x86 titan's Foundry Direct Connect 2025 event on Tuesday, SVP and GM of Intel Foundry Kevin O'Buckley teased a concept AI accelerator which will sandwich a load of speedy SRAM tiles between an AI accelerator or CPU die and a base die built on 18A-PT.