Announcement

Collapse
No announcement yet.

AVX-512 Performance Comparison: AMD Genoa vs. Intel Sapphire Rapids & Ice Lake

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • #21
    Originally posted by coder View Post
    If you normalize by core count, I think Sapphire Rapids had a lot more wins than losses. Even in absolute terms, they managed 6 or 7 wins, although 2 of those were in Open Image Denoise (which looks highly suspect) and 2 were in LeelaChess.
    how about we normalise for cost

    Xeon® Platinum 8490H $17000.00 x2 (https://ark.intel.com/content/www/us...-1-90-ghz.html)

    EPYC™ 9654 - 11,805 USD (https://www.amd.com/en/products/cpu/amd-epyc-9654)

    (we can argue all day about the amount of money people will actually pay, but that's moot here)

    and given the oh so many extra cores, EPYC wins for density. 1 Rack of EPYC will draw the same power as 1 Rack of Xeons.

    The EPYCs will cost less and get more done.

    Comment


    • #22
      Originally posted by boxie View Post

      The EPYCs will cost less and get more done.
      Not to count the performance/watt ratio: even with the worst case scenario (the Image Denoiser benchmark on page 4, where AVX512 are not activated on AMD) still reports a better power usage than Intel. In an AMD-favorable scenario (Embree, page 3) performance/ratio is consistently better.

      Comment


      • #23
        Originally posted by ms178 View Post
        What a pity that Intel couldn't make AVX-512 work on Raptor Lake. Their competitive stance against Zen 4 would be much better if their desktop SKU could also profit that much from its use. I wonder if Intel tries to compete with their HEDT lineup against the higher core count parts of Zen 4. While it came down to pricing, that would be more exciting than bringing a 6-P-core notebook chip to the desktop.
        Most consumer use cases don't need avx 512.
        it will be waste to put avx 512 on raptor lake.

        amd can't take it out from desktop ryzen because it uses same chiplet as epyc.
        I won't be surprised if the ryzen 7040 mobile series doesn't have avx 512.

        Comment


        • #24
          Originally posted by zamroni111 View Post
          Most consumer use cases don't need avx 512.
          Yes, somewhat limited applicability.

          Originally posted by zamroni111 View Post
          amd can't take it out from desktop ryzen because it uses same chiplet as epyc.
          They could've simply disabled it, which is exactly what Intel did.

          Originally posted by zamroni111 View Post
          I won't be surprised if the ryzen 7040 mobile series doesn't have avx 512.
          Given that AMD is shipping it in their desktops, I think that would be a mistake. AMD now has an ISA lead over Intel, which they shouldn't waste. Given that their implementation has good power-efficiency and probably doesn't use a whole lot of additional area, there's not much downside to carrying it forward.

          BTW, imagine when the next gen consoles launch. They'll also get it, which should probably mean it'll be used more in games.

          Comment


          • #25
            Are there any tests that show the impact of the User Interrupts in SPR?

            Comment


            • #26
              Originally posted by coder View Post
              Yes, somewhat limited applicability.


              They could've simply disabled it, which is exactly what Intel did.


              Given that AMD is shipping it in their desktops, I think that would be a mistake. AMD now has an ISA lead over Intel, which they shouldn't waste. Given that their implementation has good power-efficiency and probably doesn't use a whole lot of additional area, there's not much downside to carrying it forward.

              BTW, imagine when the next gen consoles launch. They'll also get it, which should probably mean it'll be used more in games.
              disabling it while the circuitry still exists like intel did on alder flame is waste of silicon.
              amd should either removes the circuitry from 7040 design or keeps it functional

              Comment


              • #27
                Originally posted by jayN View Post
                Are there any tests that show the impact of the User Interrupts in SPR?
                The User Interrupts support still hasn't been merged for the Linux kernel... I suppose I should write another article on that.
                Michael Larabel
                https://www.michaellarabel.com/

                Comment

                Working...
                X