“New Intel chips were often delayed and offered only small improvements over previous generations.”
As noted on HN and elsewhere, Intel’s struggle has been in manufacturing. They minimized risk (“got lazy”) when Moore’s Law reigned and their fabs were 1-2 generations ahead of everyone else, but have since squeezed every last architectural drop out of 14nm. Put Tiger Lake on TSMC 5nm and a lot of M1’s lead goes away.
The M1 is a damned fine chip and I’d like to own one. But the recent hero worship is a bit awkward to read as Apple is generally following an accepted playbook in the End Times of Moore’s law.
1. Save power by creating accelerated logic blocks without locking yourself out of algorithmic improvements.
2. Execute flawlessly. A chip respin costs 3-4 months and who knows how much money, so you need to go to production with the A0 silicon you powered up.
3. Know your software-only workloads. At best, all mathy code can be shoved down SIMD pipelines for high IPC. At worst, you’re dealing with a bunch of branchy integer code. The M1 has solved for both with vector extensions and a massive reorder buffer and register file.
It’s a great chip and Apple got there first. They’ll be rewarded with sales and increased market share, and developers will have one architecture to wrench on for all platforms.
You would be right to say that Intel couldn’t make the M1. It’s not because Intel’s fabs are struggling or Apple has found tricks no one else in the world knows about. It’s because Apple controls the entire software stack and therefore knows what they need on the chip. Intel caters to the entire market (one the M1 just noticeably shrunk) with myriad integrators, workloads, and users, and each has different concerns and priorities making accelerated algorithm blocks useless to some and not-fast-enough for others.