Apple says M5 Ultra uses a quad-die 3nm architecture, “a first for Apple silicon”, with inter-die bandwidth of 4.4TB/s+ and up to a 36-core CPU and 80-core GPU
The M6 has also arrived with all-new 2nm process technology. — With the launch of the latest Mac mini and Mac Studio computers …
Context & Ripple Effects
The quad-die claim is the fourth step in a packaging ladder Apple has been climbing since it fused two M1 Max dies into one chip back in 2022 (800GB/s of memory bandwidth) and repeated the two-die trick with 3nm silicon for the M3 Ultra in March 2025. Doubling to four dies with 4.4TB/s+ inter-die bandwidth means Apple is now scaling Ultra-tier performance by stitching dies together rather than waiting on a bigger node.
The same-day launch of the M6 on 2nm splits the lineup in two directions: cutting-edge process for mainstream Macs, aggressive packaging for the Mac Studio halo machine. The pickup was unusually broad — Bloomberg, PetaPixel and Hindustan Times alongside the usual Apple outlets — reflecting how much of the reaction framed the M5 Ultra as a local-AI box rather than a creative-workstation upgrade.
First-order effects
- Mac Studio buyers get a machine that discussion posts describe as able to run hundreds-of-billions-parameter LLMs locally, thanks to up to 512GB of unified memory and roughly 4.5x M3 Ultra's peak AI compute — with mkbhd noting a fully specced unit at $18,299.
- NVIDIA's DGX Spark loses its clearest consumer-reachable comparison point: commenters are already benchmarking the M5 Ultra's 1.2TB/s memory bandwidth at 4.4x DGX Spark's.
Second-order effects
- Windows AI-workstation rivals now have to answer a 512GB-unified-memory desktop, which pressures AMD and Intel to match Apple's memory-capacity-per-watt story or concede the local-inference niche outright.
- The price floor rises with capability: if the top configuration already sits above $18k, Apple is pulling Mac Studio away from pro-sumer territory toward rack-adjacent budgets, changing who its channel sells to.
Third-order effects
- If die-stitching keeps delivering generational leaps while node shrinks deliver the rest, multi-die packaging becomes the primary scaling lever across the industry — and Apple's UltraFusion-style interconnect a template competitors must replicate.
- Desktop towers positioned as private inference endpoints would carve out a new tier between consumer PCs and cloud GPUs, shifting some AI workloads off datacenters and making memory capacity, not FLOPS, the spec that sells machines.
The trend: Apple silicon is scaling Ultra-tier AI performance through multi-die packaging and unified-memory capacity rather than raw process-node gains, positioning high-end Macs as local inference hardware.