Ok so the thing with Intel4 and 20a is they were limited earlier versions of their respective node families that only had the libraries for a compute tile but not the entire SoC.
Well, yes although even if Meteor Lake used Intel 3 and Arrow Lake used 18A, those nodes would still only be used for the compute tile.
They were supposed to be the stepping stones into each series of nodes and instead they just turned out to be garbage.
The thing about a stepping stone is that you move off of it and onto the next node. However, when you make a product offering on a node, you have to scale up production and sustain it for years. That's much different and more costly than the internal development nodes they create.
Intel reportedly had yield problems on Intel 4 and didn't want to spend time optimizing it, because it's not their main production node. That resulted in Meteor Lake being fairly unprofitable for them. That tells me it's probably a bad idea to have a node for (a limited part of) one specific product. You want a node that's going to be high-volume, so you can justify the investments into refining it to be high-yielding and profitable.
I think this highlights a fundamental flaw in the "5 nodes in 4 years" strategy, which is that it's too expensive. Intel spent more than half a decade milking their 14nm family. They also spent about 5 years milking their 10 nm family, if you start counting all the way back from Ice Lake and include the derivative ESF node used for Raptor Lake. If you suddenly switch to a model of essentially making a new node each year, not only are you pushing the R&D teams, but you also have much less time to refine & recoup your investment on these nodes. Even if we lump together Intel 4 + 3 and 20A + 18A, to be equivalent with how I grouped those earlier nodes, it's still a more compressed timeline for recouping an investment that we're told is getting ever bigger.
BTW, I really wonder how much worse Arrow Lake would've been on Intel 3. If it truly is a 3nm-class node, then it should be mostly comparable to TSMC N3B (which was their first-gen 3nm process).
Facebook
Twitter
Instagram