Just wondering,what AMD would need to do…to at least MATCH nvidias offering in A.I/dlss/Ray tracing tech

  • bctoy@alien.topB
    link
    fedilink
    English
    arrow-up
    0
    ·
    1 year ago

    The first thing is that AMD don’t build as big chips as nvidia. In raster, RDNA3 is extremely good against Ada shader-count-wise. As for RT, you can also think that they are comparatively better at raster instead. Double the N31 GCD and you have a chip that is best in raster and between 4080/4090 in RT.

    The bigger problem is software and I doubt AMD will match nvidia there. The current AAA PT games are done with nvidia support and while it’s not nvidia-locked, it’d be great if intel/AMD optimize for it or get their own versions out.

    The path tracing updates to Portal and Cyberpunk have quite poor numbers on AMD and also on intel. Arc770 goes from being ~50% faster than 2060 to 2060 being 25% faster when you change from RT Ultra to Overdrive. This despite the intel cards’ RT hardware which is said to be much better than AMD if not at nvidia’s level.

    https://www.techpowerup.com/review/cyberpunk-2077-phantom-liberty-benchmark-test-performance-analysis/6.html

    The later path tracing updates to classic games of Serious Sam and Doom had the 6900XT close to 3070 performance. Earlier this year, I benched 6800XT vs 4090 in the old PT updated games and heavy RT games like updated Witcher3 and Cyberpunk, and 4090 was close to 3.5x of 6800XT. 7900XTX should be half of 4090’s performance then in PT like in RT heavy games.

    https://www.pcgameshardware.de/Serious-Sam-The-First-Encounter-Spiel-32399/Specials/SeSam-Ray-Traced-Benchmark-Test-1396778/2/#a1

    • ResponsibleJudge3172@alien.topB
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 year ago

      Shader count wise it’s really not a win. They stopped counting the physically present ILP shadera because of their poor scaling. Shadera don’t exist in a vacuum though, the smallest execution unit of both AMD and Nvidia are the SM/CU. And AMD currently needs more CU to match the equivalent Nvidia SM. Which was not the case last gen