The launch
On 10 November 2020 Apple introduces the M1 chip, the first SoC in the Apple Silicon line designed in-house for Mac — following the transition announcement from Intel (June 2020, WWDC). It debuts on MacBook Air, 13” MacBook Pro and Mac mini. Manufacturing process: TSMC N5 (5 nm).
The architecture
M1 integrates in a single package a CPU (4 Firestorm performance cores + 4 Icestorm efficiency cores), a GPU (up to 8 cores), a Neural Engine (16 cores for ML inference), ISP, Secure Enclave — and, architecturally most relevant, LPDDR4X RAM physically integrated on the package itself (on-package) and shared as coherent unified memory across all compute units.
There is no longer a distinction between system memory and VRAM: CPU, GPU and Neural Engine access the same addresses without explicit copies. This removes the traditional PCIe-GPU bottleneck and drastically reduces energy consumption.
Line evolution
The M1 family expands rapidly:
- M1 Pro / M1 Max (October 2021) — 14”/16” MacBook Pro, up to 64 GB unified memory
- M1 Ultra (March 2022) — Mac Studio, two M1 Max linked via UltraFusion, up to 128 GB
- M2 (June 2022), M3 (October 2023), M4 (May 2024) — successive generations with improved process and more powerful Neural Engine
- Pro/Max/Ultra variants replicated per generation
Meaning for local AI
The unified memory up to 128-192 GB in the Max/Ultra variants makes Apple workstations practical platforms for local inference of LLMs from 30 to 70 billion parameters (quantised), via:
- llama.cpp with the Metal backend (Apple’s GPU API)
- PyTorch with the
mpsdevice (Metal Performance Shaders) - MLX — dedicated AI framework released by Apple in December 2023
- Ollama, LM Studio — user-friendly runtimes based on llama.cpp
The ARM CPU + GPU + unified DRAM architectural pattern inaugurated by M1 will later be adopted by NVIDIA GB10 Grace Blackwell Superchip (2025) with explicit AI optimisation and the CUDA stack — same core idea, different silicon.
Impact
The M1 launch marked:
- The end of Apple’s 15-year Intel dependency
- The return of ARM as a mainstream desktop/laptop architecture, after Microsoft Surface ARM had had limited success
- New competitive pressure on Intel and AMD for energy-efficient designs
- The demonstration that unified memory could work at scales beyond mobile
In the Italian context
Apple Silicon has significant penetration among Italian developers, creative teams, academic researchers, professional firms and SMEs. In local AI, many Italian teams use 64-128 GB M3 Max / M4 Max MacBook Pro as LLM prototyping workstations before scaling to server infrastructure. It is one of the cases where a consumer product becomes a credible professional tool without having to move to enterprise solutions.
References: Apple M1 (10 November 2020). Apple Silicon line. TSMC N5 process (5 nm). Firestorm/Icestorm cores, Neural Engine, on-package unified memory. M1 Pro/Max/Ultra, M2/M3/M4 evolution. Frameworks: Metal, MLX, PyTorch MPS.