Apple introduces M6 and M5 Ultra
Apple’s new M6 and 512GB M5 Ultra chips are seen as a major step for on-device AI, with massive unified memory and bandwidth making serious local LLM inference feasible on a single desktop machine. Commenters weigh these capabilities against steep RAM pricing, weaker Linux support, and the economic reality that cloud-hosted frontier models often remain cheaper and faster for most workloads. The products are viewed as ideal for privacy-sensitive use cases, content creation, and well-funded enthusiasts or companies, but overkill or poor value for casual users who can rely on subscriptions and APIs instead.
Hardware & Performance
- M5 Ultra’s 512GB unified memory and 1.2 TB/s bandwidth are widely seen as impressive, especially for local AI and media work.
- Some compare memory bandwidth and compute to high‑end Nvidia GPUs (e.g., 5090, RTX 6000 Ada), noting Apple wins on RAM capacity, thermals, and simplicity, but loses on raw GPU compute and prefill speed.
- M1–M5 generation machines are often described as “still fast enough,” with several people seeing little real‑world benefit from upgrading for typical dev/creative workloads.
Local AI vs Cloud
- Many view these machines as ideal for on‑device LLMs, especially mid‑size models (Qwen 3.8 27B, Qwen 3.5 122B, etc.), with unified memory enabling larger contexts than commodity PCs.
- Others argue cloud/frontier models (Sol, Opus, etc.) are economically superior: detailed math shows $20k local rigs are hard to justify vs. $100–200/month subscriptions or cheap per‑token APIs, unless utilization is very high or strict privacy / zero‑data‑retention is required.
- There’s interest in multi‑Mac RDMA over Thunderbolt (up to 4 machines, 2TB unified RAM), but performance vs multi‑GPU Nvidia setups remains debated.
RAM, Unified Memory & Pricing
- Unified LPDDR5X RAM is praised architecturally (simple, big address space, used as “VRAM”), but upgrade prices (~$25/GB, 96→256GB = $4k, 512GB projected +$6.4k) are heavily criticized as “eye‑watering.”
- Some compare total cost favorably to building equivalent Nvidia HBM/RTX 6000 systems; others note DRAM boom, price‑fixing allegations, and expect eventual RAM price crashes.
Linux & Openness
- Major frustration: no native Linux support on newer Apple Silicon; Asahi currently tops out around M2/M3-in-progress, without full GPU acceleration.
- Users resort to Linux VMs on macOS for orchestration and tooling, but lack of GPU passthrough limits AI performance.
- Several state they’d buy Mac hardware instantly if Apple officially supported Linux or published hardware docs/drivers.
Who Buys 512GB Ultra?
- Suggested buyers: AI startups wanting on‑prem inference, firms with sensitive/regulated data, media pros (video, VFX, audio), and affluent enthusiasts.
- Many see it as a niche but valid product: cheaper and quieter than equivalent multi‑GPU racks, though clearly not “best perf/$” compared to bare Nvidia boxes.
macOS vs Linux/Windows
- Some praise macOS polish, ecosystem, and battery life; others see a long‑term decline (UI inconsistencies, Settings redesign, stability issues, CLI/tooling removal).
- A visible contingent is shifting to Linux (often with tiling WMs like Hyprland/Omarchy or KDE), citing greater customizability and better fit for dev and container‑heavy workflows.
Broader AI & Turing Test Tangents
- Side debate on whether modern LLMs have “beaten” the Turing test; opinions range from “yes, for many humans in real contexts” to “not with strong judges and long conversations.”
- Some worry AI primarily “enshittifies” the internet; others counter with global progress stats and optimistic use cases (agents running 24/7 research, personal data analysis, local private assistants).