‹ BackHN Continuity

Thread

Intelligence per Watt: Measuring Intelligence Efficiency of Local AI

169 points · 65 comments · pythonic_hell

  1. polotics · · focus · HN ↗
    Did I miss something or does this article not bother to indicate how much RAM their M4-Max had?
    1. washadjeffmad · · focus · HN ↗
      Footnote on page 3.

      >Total parameter count governs storage: at FP4, MIXTRAL-8X7B fits within 24 GB GDDR6 (Quadro RTX 6000), and GPT-OSS-120B fits within 128 GB unified memory (Apple M4 Max).

Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.