GoKawiilThe big picture: Apple announced two new chips and two refreshed desktops today. The M6 is Apple's first 2nm chip. The M5 Ultra is its first quad-die design and the most powerful silicon the company has shipped. As for the desktop PCs, there are no significant new features on either machine, but this is a real spec and pricing bump. What's changed in the past couple of years is who these boxes are for, and whether you'll be able to get one.
The new M6 chip moves to TSMC's 2nm process and is the first Apple SoC to combine all three core types at once. The CPU has 12 cores: two "super cores," four performance cores, and six efficiency cores. Super cores themselves aren't new. Apple introduced them with the M5 Pro and M5 Max in March, and M6 is the first chip to run all three tiers together.
Apple claims to have the "world's fastest" single-threaded performance, 1.2x the multithreaded throughput of the M5, and 40% faster CPU performance than the M4. The GPU moves to 12 cores with a Neural Accelerator in each one, good for nearly 30% more peak GPU compute for AI than the M5.
Rounding out the chip is a dual 16-core Neural Engine that system frameworks can dispatch to simultaneously, storage rated at twice as fast (15GB/s), and up to 170GB/s of memory bandwidth. Memory capacity caps at 32GB, which matters more than it sounds like it should.
M5 Ultra: two M5 Max chips acting as one
The Ultra is arguably the more interesting chip. Apple's UltraFusion interconnect fuses two dual-die M5 Max chips into a single quad-die package, with inter-die bandwidth above 4.4TB/s and more than 6x the connection density of the previous version. The four dies behave as one processor.
That gets you up to 36 CPU cores (12 super, 24 performance), up to 80 GPU cores, a 32-core Neural Engine, and up to 512GB of unified memory running at 1.2TB/s. Apple claims 4.5x the peak GPU compute for AI compared to the M3 Ultra and 40% faster graphics.
Note how the cheaper Mac gets the newest process node, while the flagship chip stays on the M5's 3nm process. That's normal for Ultra parts, which are assembled from Max dies and always trail a generation, but it means the M5 Ultra name undersells a chip that's a much bigger jump than the M6 in AI workloads. We've reported before that Apple doesn't plan an M6 Pro or M6 Max, skipping straight to the M7 in mid-2027.
Not for creators anymore: the local AI angle
Apple shipped macOS 26.2 last December with support for low-latency communication between Thunderbolt 5 hosts, specifically for distributed inference using MLX. Since then, hobbyists and developers have been daisy-chaining Mac minis and Mac Studios to run models far larger than any single consumer machine can hold, as an alternative to buying racks of Nvidia hardware.
... continue reading