12. November 2020 · Categories: Apple

The new Apple silicon for the Mac starts with a chip that provides roughly the same performance as the top end Intel chip in MacBooks, while consuming so little power that the new Air no longer has a fan.

The transition will likely go smoothly. As we no longer have self modifying code, translation of x64 code to ARM is now feasible at reasonable performance, and the chip has enough headroom for the losses incurred.

The new Mac mini is a missed opportunity: with the same underlying architecture as Apple TV, it would have been nice to see a home focused variant, that includes Apple TV, and serves as a home server as well as a WiFi router.

The architecture is now tightly integrated, with shared memory for all blocks. I wonder how this will translate to higher performance configurations. Either you go for really large chips to keep the integration, or you need multiple chips with corresponding problems in synchronizing memory. I suspect that large chips are now feasible, with dozens units you can plan for disabling a few to avoid defects, and the caches could be 16 way, allowing you to disable one way where you find defects.

With any multi chip solution you have limited bandwidth to address cache coherency issues, and congestion accessing the shared memory. This is why massively parallel computers put their emphasis on fast links between nodes, and not on a fast, shared memory.