An interview with Apple's Tim Millet and Tom Boger on Apple's chip strategy, making big changes to the Neural Engine after the 2017 transformer paper, and more
'No other platform can touch our power performance per watt. That's the tangible benefit to users,' Tom Boger, Vice President of Mac Product Marketing at Apple, said.
Context & Ripple Effects
Apple’s chip story has been framed as a long-running transition: the M1 design discussion centered on unified memory architecture, followed by coverage of M2 and the move away from Intel in Macs.
Later coverage connected Apple silicon to renewed Mac gaming efforts and features such as Dynamic Caching and Game Mode. This interview adds AI-oriented architecture to that same custom-silicon narrative.
First-order effects
- Apple publicly ties major Neural Engine changes to the transformer-era shift in AI workloads, making that accelerator a more explicit part of its chip strategy.
- Apple’s performance-per-watt claim reinforces efficiency as the user-facing measure it wants associated with its integrated platform; the interview does not provide an independent comparison.
Second-order effects
- The emphasis raises the importance of how well competing platforms balance specialized AI compute with overall power efficiency, rather than treating AI acceleration as a standalone specification.
- For Mac software and game developers already benefiting from Apple-silicon features, the message is that workload-specific hardware capabilities remain central to platform differentiation.
Third-order effects
- If this design approach persists, AI capability will be differentiated increasingly at the system-architecture level—memory, compute blocks, and software integration—not merely by adding a generic accelerator.
- That favors firms able to sustain long chip-development cycles and tightly align silicon with platform software, an advantage suggested by Apple’s earlier M2 and Intel-transition strategy discussion.
The trend: The story is one data point in the shift toward vertically integrated, heterogeneous AI hardware designed around efficiency and specific workloads.