Apple Announces A19 and A19 Pro SoCs (morethanmoore.substack.com)

Apple unveiled the A19 and A19 Pro SoCs alongside the iPhone 17 lineup, migrating to TSMC’s higher‑performance 3nm N3P node (an optical shrink of N3E) for modest density and power/performance gains. Both chips keep Apple’s familiar 2+4 CPU core layout; the A19 Pro adds a 50% larger last‑level cache (from 24MB to 36MB) and performance‑core front‑end and branch‑prediction improvements, with Apple claiming up to 20% CPU gains vs. iPhone 15 Pro and up to 40% sustained performance vs. iPhone 16 Pro (partly thanks to thermal/cooling changes). GPU-wise the new Apple10 architecture ships with 5 GPU cores in A19 and 6 in A19 Pro, introduces tensor cores, doubles FP16 throughput relative to prior generations, and refines dynamic caching and “unified image compression” to reduce memory/bandwidth pressure. For the AI/ML community these changes matter: doubled FP16 throughput plus tensor cores and better GPU memory handling directly boost on‑device ML throughput and energy efficiency—enabling faster inference, more aggressive quantized models, and lower latency for neural workloads. Equally important is the hardware security advance: Apple’s “Memory Integrity Enforcement” implements Arm’s Enhanced Memory Tagging Extension (EMTE) plus microarchitectural mitigations to prevent tag‑leakage via speculation. That combination both hardens iOS against buffer‑overflow/use‑after‑free exploits (critical for protecting models and data) and gives developers new primitives to build safer, more reliable ML applications on device.