3 ms·

Doubling floating point performance is disappointing?

• zamadatix 2y ago
  Are you referring to "up to doubling AVX512 SIMD performance"? Floating point performance certainly didn't double in total, that SIMD includes non float instructions too, and even in heavily AVX512 workloads the actual performance increase is rarely anywhere close to double because unfused AVX512 ops are rarely anywhere near the performance of just doing AVX256 ops. https://i.imgur.com/Y1pJFq5.png https://i.imgur.com/Y1pJFq5.png This type of workload increased 27%, nearly half the average across the board performance increase Phoronix found for these types of task for Zen 3 -> Zen 4. https://i.imgur.com/xaD3x9Z.png https://i.imgur.com/xaD3x9Z.png
  • vardump 2y ago
    If you have a suitable workload, you do get nearly 100% improvement. There are 2x FPU resources now. Obviously less with generic code. Amdahl’s law and all.
    • zamadatix 2y ago
      There are certainly not 2x the FPU resources now, there are up to 2x of certain FPU resources for certain instructions.