ISSCC keynote: Apple silicon SVP outlines neural engine roadmap to 2029, 'memory-first' architecture and 2nm transition
Apple's silicon roadmap outruns its advanced packaging base
In a keynote at the International Solid-State Circuits Conference, Apple's SVP of Silicon Technologies, Priya Raghunathan, described a neural engine roadmap extending through 2029, including a transition to a 2-nanometer class process and an architecture she characterized as 'memory-first.' She said Apple now treats model weights as a tiered memory problem, with on-package SRAM and new compression accelerators, and that the current generation sustains double-digit tokens per second on a 3-billion-parameter model inside a phone thermal envelope. The next node, she said, targets 'four times that memory budget.' Raghunathan declined to name foundry partners and would not discuss tooling lead times. One slide showed a packaging roadmap with two sourcing paths through 2028, which she described as 'resilience by design.' Asked about the 2GB developer residency floor, she said it was 'a quality statement, not a hardware gate,' to audience laughter.
The roadmap is now public to 2029 with a memory-first architecture and a 2nm-class transition. The load-bearing detail is the packaging slide: two sourcing paths through 2028, stated as design intent rather than a proud supply-chain claim.
Two sourcing paths through 2028 is a message to the supply chain, delivered on a conference slide. And the room laughed at 'a quality statement, not a hardware gate' because the room knows a gate when it sees one.
- Roadmap extends through 2029 with a 2-nanometer class process and 'memory-first' design
- Current generation sustains double-digit tokens per second on a 3-billion-parameter model; next node targets 4x memory budget
- Packaging roadmap shows two sourcing paths through 2028
- SVP declined to name foundry partners or discuss tooling lead times