Cat_Daddy [any, any]

  • 1 Post
  • 42 Comments
Joined 9 days ago
cake
Cake day: July 30th, 2025

help-circle






  • I haven’t been in the industry for a while, but in a past life I did a lot of computational science. Intel’s Fortran compiler—and by extension their chips’ abilities to run Fortran calculations—was light-years ahead of the nearest competition. We had an AMD-based supercomputer (a bit older design, but still relatively new at the time; I think it was about a year old at this point) that was slower than a new Intel-based desktop cluster I had just finished building. The cluster was about a 4th the specs of the supercomputer across the board, mind you; it was really just designed to be a terminal, but it still had to be able to run the computing software so researchers could set up the calculations before moving the files to a supercomputer for calculating, or if they wanted a quick one-off computation, for example.

    So as a joke I ran one of my big calculations on it just to stress test it and make sure I’d done the heat sinks right. I fully planned to kill the job once the machine warmed up, but it finished before I could kill it. I ran several of these “drag races” between the new computer and multiple of our supercomputers and clusters (all AMD of varying age) and none of them were nearly as fast as this new machine. The supercomputer I mentioned earlier even had then-state-of-the-art PCIe solid state drives, because I/O in computational science is frequently the largest bottleneck, but the new cluster just had a standard hard drive. And even still the Intel machine was faster. Meaning if we’d had a 1-to-1 comparison the differences would have been even more drastic. That speedup comes down to the tight coupling between their Fortran compiler and their assembly language. They could make use of, for example, a matrix transpose call as a single instruction that didn’t exist in the AMD assembly, and so took several instructions. And since computational science is almost totally linear algebra, this leads to a huge jump in computing speed.