Integer addition algorithm could reduce energy needs of AI by 95%

Comments

• technocrit@lemmy.dbzer0.com ⁨2⁩ ⁨months⁩ ago

  This is an extremely misleading headline. From the abstract:

  … applying the L-Mul operation in tensor processing hardware can potentially reduce 95% energy cost by element-wise floating point tensor multiplications and 80% energy cost of dot products.

  In other words… This method of computation could save 95% of the energy spent on floating point multiplication (and 80% of dot products)… Not 95% of total energy.

  source

  • AndrasKrigare@beehaw.org ⁨2⁩ ⁨months⁩ ago

    I’d say it’s not just misleading but incorrect if it says “integer” but it’s actually floats.

    source
  • IrritableOcelot@beehaw.org ⁨2⁩ ⁨months⁩ ago

    Good point. Though, the vast majority of ML training and use is tensor math on floating points, so largely dot and cross products, among other matrix operations.

    source
• django@discuss.tchncs.de ⁨2⁩ ⁨months⁩ ago

  Rebound effect getting ready behind a tree

  source

  • halm@leminal.space ⁨2⁩ ⁨months⁩ ago

    Came here to say that a 95% reduction in energy consumption will only greenwash a corresponding or larger increase in usage — but yours is of course the correct response! 👏👏👏

    source
  • skarn@discuss.tchncs.de ⁨2⁩ ⁨months⁩ ago

    But… Isn’t that kind of the point? Slashing computational cost so that we can deploy that stuff wherever it’s needed without a tenfold increase in the world’s energy bill?

    Whether we should do that at all is a very different question.

    source

    • halm@leminal.space ⁨2⁩ ⁨months⁩ ago

      “Wherever it’s needed” is the operative term here, isn’t it? Looking at how it’s already being implemented, nobody seems to bother asking whether “AI” is really needed.

      source

      • -> View More Comments
    • Artyom@lemm.ee ⁨2⁩ ⁨months⁩ ago

      Improving the technology behind AI will only increase the return on investment per watt, so you’ll want to spend even more on it than before. This would more than likely increase the energy demands (assuming it doesn’t turn into vaporware).

      source
• Jayjader@jlai.lu ⁨2⁩ ⁨months⁩ ago

  I wonder what other applications this might have outside of machine learning. I don’t know if, for example, intensive 3d games absolutely need 16bit floats (or larger), or if it would make sense to try using this “additive implementation” for their floating point multiplicative as well. Modern desktop gaming PCs can easily slurp up to 800W.

  source

  • DdCno1@beehaw.org ⁨2⁩ ⁨months⁩ ago

    I have a vague memory of a 3D engine from the 1990s using an approach like this, but I’m not entirely sure.

    source

    • IrritableOcelot@beehaw.org ⁨2⁩ ⁨months⁩ ago

      I think you’re thinking of the famous fast inverse square root algorithm from Quake.

      With respect to the top comment, the only reason 3d graphics are possible (even at 850W of power consumption) is due to taking a bunch of shortcuts and approximations like culling of polygons. If its a reasonable shortcut it either has or will be taken.

      source