The M2 is more advanced than it seemed

Original Article ↗ Hacker News Discussion ↗

Apple’s M2 chip is highlighted as a bigger architectural step than it first appeared, notably by adding ARMv8.6-A features like bfloat16 that improve on-device AI and future client-side model training, where the M1 may increasingly lag. Commenters weigh this against practical factors such as real-world macOS responsiveness, RAM and storage constraints, corporate security tooling overhead, and long upgrade cycles, while also comparing Apple Silicon to AMD, Qualcomm, and x86 laptops on performance, power efficiency, and battery life. There is broad agreement that hardware is advancing quickly, but that bloated apps, weak ML tooling on macOS, and uncertain economics for on-device vs cloud AI will determine how much of that potential users actually see.

M2, ARMv8.6-A, and bfloat16

  • Key technical point: M2 moves from ARMv8.5-A to ARMv8.6-A, adding bfloat16 support and AI-relevant instructions.
  • Commenters note this mainly benefits training and heavy numerical workloads; inference often uses more aggressive quantization.

Training vs inference and device longevity

  • Several posts say bfloat16 is more relevant to training than to typical on-device inference.
  • Some expect more on-device training/fine‑tuning on personal data (e.g., grammar, personalization), making older chips less ideal.
  • Others argue by the time that matters, both M1 and M2 will be “old,” and most people will upgrade to later generations anyway.
  • Long‑cycle users (8–10 years) push back that future ML support does matter to them.

PC laptop competition and CPU trade‑offs

  • Discussion about whether AMD/Intel/Qualcomm laptops can match Apple’s perf/W:
    • AMD Zen 4 mobile is seen as roughly competitive in compute, with separate AI blocks and bfloat16 too.
    • Apple is viewed as prioritizing high IPC and tight integration; x86 vendors prioritize cheaper, smaller cores and flexibility.
  • Qualcomm’s upcoming Snapdragon X Elite / Oryon is repeatedly mentioned as an ARM challenger but with open questions on power use and real delivery.

Battery life and real‑world performance

  • Some say AMD laptops can match or exceed Apple on raw performance and even browsing battery life, depending on tests and battery size.
  • Others buy Macs specifically for 18–20h realistic battery life and feel non‑Apple machines still lag in that dimension.
  • Thread notes big differences in measurement methodology and that Apple’s software stack is highly optimized for idle and low‑load power.

Software bloat, UI frameworks, and responsiveness

  • Many complaints target Slack and Electron‑style apps that feel laggy even on M2.
  • Debate over whether JavaScript is inherently too slow, or whether poor engineering and bloated frameworks are the real issue.
  • Some argue cross‑platform browser‑based UIs are economically irresistible but guarantee mediocre UX; others point to well‑optimized JS (e.g., GNOME) as proof that JS itself isn’t the problem.
  • Multiple attempts at native, declarative, cross‑platform UI toolkits (Qt/QML, Flutter, Slint, etc.) are discussed; no clear consensus “winner.”

On‑device AI vs cloud and local LLM use

  • One view: rapid hardware advances and heterogeneity will push most advanced AI into the cloud, with only minimal on‑device work.
  • Counter‑view: economics of cloud GPU costs strongly incentivize pushing as much inference as possible to edge devices, requiring hybrid frameworks.
  • Several users report good experiences running local LLMs and image models on M1/M2 (especially M2 Pro with high bandwidth), using various mac‑focused toolchains.
  • Concerns raised about stability and thermals on fanless MacBooks under sustained LLM or OBS loads; others report stable experiences when staying within RAM limits and avoiding locking large models into memory.

macOS lag spikes and corporate tooling

  • A number of reports of macOS UI “lag spikes” on M1/M2:
    • Often correlated with low RAM (8 GB) and aggressive swapping.
    • Corporate security/management suites (JAMF, CrowdStrike, etc.) and aggressive file scanning are repeatedly blamed for severe slowdowns.
  • Other users with lightly loaded or personal Macs say they see nearly no stutters, suggesting environment and tooling matter more than silicon.

Upgrade and trade‑in considerations

  • Some users debate upgrading from M1 to M2/M3 vs waiting for M3 Air.
  • Advice trends toward:
    • Prioritizing RAM and SSD capacity over small CPU generational gains, especially for heavy multitasking/containers.
    • Watching out for slower low‑capacity SSDs on newer base models.
  • Experiences with Apple trade‑in vary:
    • Online quotes can be “decent”; in‑store assessments can be stricter.
    • Others report smooth mail‑in trade‑ins and emphasize you can always reject an adjusted offer and keep the device.

Technical side‑notes

  • Brief technical digression on whether bfloat16 can be emulated via float16: consensus is that ranges differ, so you can store but not safely compute that way; converting via float32 is suggested.
  • One commenter notes benchmark data implying M2→M3 single‑thread gains are larger than M1→M2, slightly contradicting descriptions of M2 as only “evolutionary.”