you are viewing a single comment's thread
view the rest of the comments
[–] 5 points 12 hours ago (4 children)

The HBM (slop memory) doesn’t work in a real computer.

They’ll have to just get rid of it because it’s literally useless for anything other than slop generation.

So regardless of what our dumbass market dictates, that HBM is useless, full stop.

  • source
  • parent
  • hideshow 4 child comments
  • [–] 3 points 10 hours ago* (last edited 10 hours ago) (1 child)

    True. I think that's about a quarter of the dram produced. Though it could be used in public sector research and not just thrown away.

    Astronomy and Biology research benefit from AI image analysis models. Because it's not just generating slop text. It's used to analyze image data to find potential points of interest. Reducing an impossible number of raw data points (stars, cells, etc) down to a manageable number that can be analyzed and confirmed by an actual human.

    This isn't the "AI" people talk about when they say "AI". It was just called "machine learning with computer vision models" before all the AI marketing crap.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 1 point 6 hours ago*

    I don’t know enough about the mechanical workings of things like radiological image recognition to know whether or not HBM would be useful for them, but hopefully you’re right.

    I believe HBM makes up more than a quarter of current production, with all 3 major memory manufacturers prioritizing HBM.

  • source
  • parent
  • [–] 2 points 10 hours ago* (last edited 10 hours ago) (1 child)

    HBM is far from useless - it's a really good, compact memory architecture with tons of bandwidth. Find a way to adapt it to DRAM modules and we're gold.

    Worst case we get GPUs that can play AAAA games at 8k 120fps

  • source
  • parent
  • hideshow 1 child comment
  • [–] 2 points 5 hours ago

    The GPUs designed for inference (slop) are not the same kind of GPUs that run video games either.

    Here’s an article explaining why HBM is really difficult to integrate into a real computer, and why doing so would eliminate nearly all of the advantages of using HBM (while failing to eliminate the things that make HBM much more expensive to both produce and power than regular DRAM).

    Your proposed “worst case” scenario would be far better than reality’s “best case” scenario.

    (The real best case scenario would have been to not build all this worthless bullshit in the first place, but its too late for that).

  • source
  • parent