Apple M5 chip

Original Article ↗ Hacker News Discussion ↗

LLM Performance, Memory Bandwidth & Capacity

  • Much debate centers on whether the base M5’s 153 GB/s unified memory bandwidth and 32 GB max RAM are enough for “proper” local LLM use.
  • Several comments explain that for large models every parameter effectively must be touched per token, making memory bandwidth and capacity the main bottlenecks; insufficient bandwidth caps tokens/sec regardless of compute.
  • Others argue the base M5 is fine for smaller or 4-bit quantized models and that Pro/Max/Ultra variants (projected ~600+ GB/s) plus 128–512 GB unified RAM will be the real LLM workhorses.
  • There’s disagreement over cost‑effectiveness versus high-end GPUs (e.g., 4090/DGX Spark): Apple wins on power and noise, loses on peak bandwidth.

Apple’s “AI” Branding and webAI Callout

  • Several note Apple previously avoided the generic “AI” label in favor of “machine learning” or “Apple Intelligence”, but the M5 press uses “AI” heavily.
  • The explicit mention of webAI as an example of on‑device LLM usage is seen as a mutually beneficial showcase of Apple’s local‑first strategy and a smaller partner that isn’t Meta/OpenAI/Chinese.

Hardware Outpaces Software

  • Strong consensus that Apple’s silicon and devices are outstanding, while macOS/iOS quality and UX have regressed.
  • Many describe Tahoe/iOS 26 as sluggish and glitchy, with laggy animations and battery drain; some tie this to an Electron private‑API bug, others to general bloat and “iOS‑ification”.
  • Long subthread rehashes whether Apple is fundamentally a hardware, software, or “systems/product” company, using revenue splits, historic quotes, and comparisons to Windows/Linux UX.

Gaming on Mac

  • Multiple users report great raw performance (e.g., Death Stranding, Cyberpunk, WoW) via native ports or GPTK/Wine, but overall gaming is “brittle”: anti‑cheat, Steam limitations, APIs, and OS instability across versions.
  • Explanations: small Mac gamer market, constant API/compat breaks (32‑bit removal, OpenGL deprecation, signing changes), Metal’s isolation vs DirectX/Vulkan, and Apple’s limited incentives (no cut on Steam).

Linux, Asahi & Openness

  • Strong desire for M‑series hardware with first‑class Linux support (or even an Apple‑sold Linux/Windows line).
  • Asahi is praised on M1/M2 but stalled on newer chips; missing sleep, TB/DP, video codecs, and full power management make it non‑mainstream.
  • Some say macOS still feels “dev‑hostile” versus Linux in openness, scripting, and window management despite its Unix base.

M5 Lineup, Specs & Marketing Claims

  • Confusion that only a base M5 exists so far, in a 14" MacBook Pro and iPad Pro, with no 16", Mac mini, or Pro/Max/Ultra yet; most expect higher‑end parts in the next cycle.
  • Bandwidth figure (1,224 Gbps = 153 GB/s) is seen as good for a base SoC but unimpressive versus earlier Max/Ultra parts and discrete GPUs.
  • Several question Apple’s “up to 3.5x” and “4x AI GPU” claims, noting that real‑world examples in the press releases mostly show ~1.2–2.3x improvements.

Neural Engine, “Neural Accelerators” & Software Stack

  • Discussion parses Apple’s multiple matmul paths: SIMD/AMX/SME on CPU, GPU tensor-style units, and the Neural Engine (ANE).
  • Some think the new “neural accelerators” are GPU tensor arrays; others highlight that ANE remains optimized for low‑power CNN‑style inference and is awkward to target (CoreML/ONNX only).
  • Developers complain that Apple’s ML stack (Metal, CoreML, MLX, MPS) is powerful but fragmented and less aligned with mainstream PyTorch/CUDA ecosystems.

AI Strategy, Energy Use & Climate Concerns

  • One camp sees Apple’s local‑first AI as a niche “persona/photo tricks” sideshow that’s dragging OS quality and wasting money; others argue on‑device private AI plus strong silicon is a long‑term differentiator.
  • Some refuse to use Apple Intelligence for climate reasons; others counter with claims that inference energy per query is small compared to overall personal footprint, though training remains energy‑intensive.
  • There’s skepticism that Apple can avoid a “Siri vs everyone else” repeat if its local models lag far behind cloud SOTA.