‹ BackHN Continuity

Thread

Micron CEO Says Memory Supply Will Be Much Tighter in 2027 and 2028 Than in 2026

394 points · 443 comments · speckx

  1. zsoltkacsandi · · focus · HN ↗
    No one will use AI products if there won't be consumer products because of RAM shortage.
    1. pickleglitch · · focus · HN ↗
      The plan is to make us all use shitty hardware like a chromebook and rent compute from the cloud for anything heavier than a web browser.
      1. cube00 · · focus · HN ↗
        > anything heavier than a web browser

        The way Chrome feasts on RAM, might need the cloud for that too.

        1. KellyCriterion · · focus · HN ↗
          Honestly, FF became similar worse: On my 16 GB machine, when surfing some hours FF has allocated around 10-11 GB and for whatever reason it does not free them, unless I exit and restart it
          1. StillBored · · focus · HN ↗
            On linux do something like `systemd-run --user --scope -p MemoryHigh=XG -p MemoryMax=X+2G firefox` to constrain it since the firefox people refuse to implement the memory limits logic that say Edge has..

            But at the same time, this isn't just a Firefox issue, although they may be worse than most. Its a fundamental tenant of GC languages. You trade CPU (for the garbage collection) for RAM utilization. IIRC This was apparent by the late 70s with Dijkstra, but very clearly indicated in say Hertz where they are explicitly scaling GC overhead vs memory footprint to find that it takes ~5x the ram to approach the efficiency of an explicit allocate/deallocate model. Lower footprints are achievable but the GC overhead goes way up just to reach 2x the ram footprint such that the GC tends to dominate the CPU utilization.

            And there are lots of papers like this for long running processes. GC works well if your model is basically allocate as much as you like, exit the process before actually needing to do a proper GC pass. Especially for contended systems where the memory is returned to the OS. This is why say PHP tended not to have this problem, because the memory was largely freed at the end of every request.

Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.