Redfire75369’s avatarRedfire75369’s Twitter Archive—№ 1,365

          1. Analysis of Alchemist's Performance/Area (P/A). There are iso-node and node-ignorant comparisons to Maxwell, Pascal, Turing, Ampere, Polaris, Vega, Navi and Navi II. Some notes before we start off, Alchemist likely has very good transcoding for its performance-tier (..)
            Permalink On twitter.com Twitter logo Redfire75369’s avatar ♻️ 1 Retweets ❤️ 12 Favorites Mood +3 🙂
        1. …in reply to @Redfire75369
          (..) (Quicksync already does this) which may add some die bulk. Intel is dedicating a large portion of its die space to matrix engines, which means it will likely outperform Nvidia's GPUs in matrix math. However, this adds significant die bulk.
          Permalink On twitter.com Twitter logo Redfire75369’s avatar ❤️ 5 Favorites Mood -8 🙁
      1. …in reply to @Redfire75369
        We'll start off with the 32 XC die, although both dies will be present in this analysis. Performance: RTX 3070 (Cut-down GA104) / RX 6700XT (Navi 22) Die Size: 379 mm² (TSMC N6) GA104: 392 mm² (SS 8LPE) [RTX 3070 Ti] TU104: 545 mm² (TSMC N12) [RTX 2080]
        Permalink On twitter.com Twitter logo Redfire75369’s avatar ❤️ 6 Favorites Mood -6 🙁
    1. …in reply to @Redfire75369
      Navi 22: 335 mm² (TSMC N7P) [RX 6700XT] Navi 10: 251 mm² (TSMC N7P) [RX 5700XT] Vega 20: 331 mm² (TSMC N7) [Radeon VII] Iso-node: P/A is worse than Ampere and Turing. P/A is worse than Navi II, possibly worse than Navi, and likely better than Vega.
      Permalink On twitter.com Twitter logo Redfire75369’s avatar ❤️ 4 Favorites Mood -7 🙁
  1. …in reply to @Redfire75369
    Node-ignorant: P/A is approximately equal to Ampere, and much better than Turing. P/A is worse than Navi II, slightly better than Navi and better than Vega.
    On twitter.com Twitter logo Redfire75369’s avatar ❤️ 4 Favorites Mood +3 🙂
    1. …in reply to @Redfire75369
      Next up, 8 XC die. In general, this will have worse P/A characteristics due to fixed components on all sizes. Performance: GTX 1650 (Cut-down TU117) Die Size: 153 mm² (TSMC N6)
      Permalink On twitter.com Twitter logo Redfire75369’s avatar ❤️ 4 Favorites Mood -9 🙁
      1. …in reply to @Redfire75369
        GA107: Unkown (SS 8LP#) [RTX 3050 Ti] TU117: 200 mm² (TSMC N12) [GTX 1650] GP107: 132 mm² (SS 14LP#) [GTX 1050 Ti] GM107: 148 mm² (TSMC 28nm) [GTX 750 Ti] Navi 24: Unknown (TSMC N6) [RX 6500M] Navi 14: 158 mm² (TSMC N7P) [RX 5500XT] Polaris 20: 232 mm² (GF 14HP) [RX 580]
        Permalink On twitter.com Twitter logo Redfire75369’s avatar ❤️ 4 Favorites Mood 0
        1. …in reply to @Redfire75369
          Iso-node: P/A is significantly worse than Turing, Pascal, and Maxwell. P/A is significantly worse than Navi and Polaris. Node-ignorant: P/A is better than Turing, Pascal and Maxwell. P/A is worse than Navi and better than Polaris.
          Permalink On twitter.com Twitter logo Redfire75369’s avatar ❤️ 3 Favorites Mood -5 🙁
          1. …in reply to @Redfire75369
            This is not the main focus of this analysis, but here's a short performance/watt (P/W) analysis of the larger die. Alchemist 32 XC - 225-235W RTX 3070 - 220W RTX 2080 Ti - 250W RX 6700XT - 230W P/W is worse than Ampere, Turing and Navi II (at peak V/f).I
            Permalink On twitter.com Twitter logo Redfire75369’s avatar ❤️ 4 Favorites Mood -6 🙁
            1. …in reply to @Redfire75369
              It's probably better than Ampere in a more reasonable point on the V/f curve, but still worse than Navi II.
              Permalink On twitter.com Twitter logo Redfire75369’s avatar ❤️ 6 Favorites Mood -1 🙁