Women in Technology

Hear us Roar



Article:
  Confessions of the World's Largest Switcher
Subject:   Flop calculations are off
Date:   2003-10-30 14:52:49
From:   anonymous2
In addition to the two floating point units, there's also the vector unit which can do 4 * 32-bit floating point multiply accumulates (vmaddfp) per cycle. I'd have to dig deeper into the docs to find if the bus can fully supply all three units without stalls.
Full Threads Oldest First

Showing messages 1 through 2 of 2.

  • Flop calculations are off
    2003-10-30 17:35:09  anonymous2 [View]

    Yes, but those are single-precision calculations. As such they probably won't be that useful to most people doing high-performance computing.
    • Flop calculations are off
      2003-10-30 23:00:42  anonymous2 [View]

      Its a matter of algorithm design, if you can split an algorithm into two single precisions streams, and then just do the final combine in double precision you can get some mighty mighty mighty impressive performance from a G5

      Often double precision is only necessary because two inputs have massively different scales. If you are working with similarly scaled numbers then you can get away single precision... and then just use the double precision for recombining the various scales.