New Framework Desktop Option with AMD Ryzen AI Max+ Pro 495 and 192GB Memory

Original Article ↗ Hacker News Discussion ↗

A new Framework Desktop configuration with an AMD Ryzen AI Max+ 495 chip and 192GB of LPDDR5X unified memory is drawing interest from people who want a compact, energy‑efficient box for local LLM and AI workloads, but also concern over price, availability, and long‑term RAM shortages. Commenters note that while the huge memory capacity enables larger or multiple models, relatively modest memory bandwidth limits performance on very large dense LLMs, making mixture‑of‑experts and smaller models more practical. There is also debate over Framework’s choice of soldered, non‑ECC unified memory and a proprietary SoC versus a more repairable, socketed design, highlighting a trade‑off between modularity and high‑bandwidth AI‑oriented architectures.

RAM capacity, pricing & market outlook

  • Existing 128GB Framework Desktop configs have become much more expensive and often out of stock, widely attributed to a DRAM shortage.
  • Company responses say there’s no sign of LPDDR5X prices dropping in the next year; expectations are for substantial increases over the next six months.
  • Shared market research cited predicts DRAM prices may not normalize until ~2030–2031, even with new Chinese capacity.
  • Past RAM glut (e.g., 2023 losses at Micron) reduced investment; export controls on fab tools are seen as further constraining supply.
  • Some argue memory makers have little incentive to boost output, calling this a market failure where governments should intervene.
  • Roadmaps mentioned: future “Medusa Halo” with 384‑bit LPDDR6, targeting 256GB at launch and potentially 512GB later, but timelines (≈1.5 years) and vendor roadmaps are treated as uncertain.

Performance, LLMs & real‑world use

  • Many buyers use the current 395 128GB system as a local LLM or dev server and are generally happy.
  • Several note memory bandwidth, not capacity, is the main bottleneck: ~256 GB/s on 395 vs ~273 GB/s on the new 495 is called a modest uplift.
  • Experience reports:
    • 35B MoE models (e.g., Qwen variants) can reach acceptable speeds; dense 27B is already at the lower tolerance limit, and 70B dense is seen as impractical.
    • Others report good performance with quantized larger models and multiple concurrent agents when configured carefully.
  • A recurring theme: very large RAM (192GB+) is most useful for running multiple smaller/mid‑sized models simultaneously, or mixing LLMs with image/video models, rather than a single huge dense model.
  • Some use the machine for 1440p/4K gaming with upscaling and are satisfied, but serious gamers are advised to consider a discrete GPU build.

Design, repairability & form factor

  • The motherboard uses standard Mini‑ITX and the PSU is FlexATX; several people have put the board into standard ATX cases.
  • Heavy debate over soldered unified memory:
    • One side argues it undermines Framework’s repairability ethos and that a socketed CPU/RAM desktop would be more on‑brand and sufficient for most users.
    • The counterargument: unified high‑bandwidth memory for large local AI workloads essentially requires soldered LPDDR/HBM due to signal‑integrity limits; socketable alternatives can’t match the same bandwidth/AI use case, and generic repairable desktops already exist.

Naming, positioning & competitors

  • The “Ryzen AI Max+ Pro 495” name is widely mocked as bloated and confusing; others defend it as systematic (AI = NPU, Max+ = memory channels, Pro = security features).
  • Bandwidth comparisons to rumored Apple M‑series Ultra chips highlight that Framework’s system targets a lower‑cost but also lower‑bandwidth niche.
  • Concerns are raised that high demand (similar to the 395) may keep the new 192GB SKU expensive and hard to buy.

ECC & reliability

  • Some dismiss the product over non‑ECC memory; others counter that ECC does report and correct errors when supported end‑to‑end and remains valuable for detecting failing RAM.
  • A side remark contrasts the rarity of RAM errors with the ubiquity of LLM hallucinations, but this is rejected as missing the point about hardware reliability.

Website UX & accessibility

  • A user reports keyboard‑navigation and screen‑reader issues with the email signup form.
  • The company acknowledges it as a bug and adds it to their backlog, which is positively received.