Getting 50 GB/S Back from the Apple Neural Engine

(eiln.github.io)

41 points | by eiln 2 days ago

6 comments

  • VladVladikoff 48 minutes ago
    This website hijacked my back button during a simple page load. You should fix that, it’s not an acceptable way to behave.
    • switchbak 26 minutes ago
      I’m shocked that’s still possible, shouldn’t we be able to fix that at the browser level?
      • busymom0 23 minutes ago
        Just like how browsers disable or mute auto playing videos, they should be able to prevent back button hijacking until user has clicking a link or button.
        • jchw 7 minutes ago
          Some websites like PC Gamer have some shitty middleware (that seems to be very careful to hide its identity, only appear once per session, and not immediately register on page load, probably to try to delay browsers banning the ever loving fuck out of their practices) that uses pushState to hijack the back button to try to SPAM more articles at you, probably in hopes of grabbing your attention. I would be in favor of permanently banning domains that do this even once (intentionally and not just getting hacked) from ever executing JavaScript ever again and forcing them to switch domains and lose their SEO.
  • bee_rider 25 minutes ago
    Nice investigation.

    It is always surprising to me when a nice round number like 1MiB results in the “bad performance” configuration (although it happens).

    Are you sure erratum is the right word in this context? I usually see it used to describe the notice that a document has an error in it.

  • thenewwazoo 13 minutes ago
    "Apparently the memory controller's throughput has a dominant harmonic with wavelength 2048 in tensor-dimension space."

    That got a laugh out of me.

  • Neywiny 1 hour ago
    Just checking here- this systemverilog is a hypothetical telling of what you think is going on? Or do you have the actual source of the RTL?
  • eiln 2 days ago
    RTL performance erratum in the Apple M3 Neural Engine throttles DRAM weight streaming throughput down to 17–19 GB/s from the nominal 45–60 GB/s. Avoiding the problematic path in the kernel DMA engine's speculative prefetch ring increased Llama 3.2 1B token throughput from 10.0 to 24.3 tokens/s.
  • RantyDave 27 minutes ago
    Ummm, wow. That's really bad.