Intel Arc B580 Delivers Promising Linux GPU Compute Potential For Battlemage

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts
  • sobrus
    Senior Member
    • Apr 2021
    • 191

    #11
    As far as I know RDNA3 has/had some issues too.

    The main difference vs RDNA2 is that it is dual issue architecture, but it's not always working. And it can't dual issue packed math - so in reduced precision workloads its peak performance is basically halved. Sadly, we don't have clpeak results for half precision here.
    It wouldn't be a problem if the cards were not nerfed vs RDNA2 - for example we have 4608 cores in 6800XT vs 3840 cores in 7800XT.

    And it seems to show somewhat, I ran clpeak on my undervolted 6800xt (2.2Ghz max)
    - ~19814 in single precision,
    - ~38800 half precision
    - ~3860 in integer compute
    - not to mention ~1225 double precision (RDNA2 has 1:16 FP64 instead of 1:32).

    It reaches almost peak theoretical performance (4608 * 2200mhz * 2 = ~20Tflops). Since it beats higher clocked 7800XT. I guess there are still issues with RDNA3 dual issuing Even for FP32.

    Having a look here:
    Editor’s Note (6/14/2023): We have a new article that reevaluates the cache latency of Navi 31, so please refer to that article for some new latency data.

    we can see that RDNA3 is faster only for FP32.
    I wonder if there is performance difference when running XeSS (INT8) on RDNA2 vs RDNA3 if 6900XT can be 30% faster than 7900XTX.

    Please take it with a grain of salt, as I am not an GPU expert in any way But these RDNA3 results are way below their advertised performance.
    7800XT should push 37 TFLOPS with FP32, not 16.5.

    edit:
    As a sidenote, I wonder what they are doing.
    - they have CDNA which is strong at compute and with AI (WMMA instructions) but not for consumers
    - and RDNA which doesn't have WMMA and only does FP16 (2x faster than FP32)
    - and RDNA2 which doesn't have WMMA but is strong at general compute and rapid packed math down to INT4 (8x faster than FP32)
    - and RDNA3 which is poor at packed math but has WMMA for AI.
    - and RDNA4 which again will be something different.

    I wonder how they could do FSR using AI if every single their GPU generation has different compute properties?
    Last edited by sobrus; 12 December 2024, 04:17 PM.

    Comment

    • darkbasic
      Senior Member
      • Nov 2009
      • 3085

      #12
      Does it support SR-IOV?
      ## VGA ##
      AMD: X1950XTX, HD3870, HD5870
      Intel: GMA45, HD3000 (Core i5 2500K)

      Comment

      • DiamondAngle
        Junior Member
        • Oct 2017
        • 46

        #13
        Originally posted by sobrus View Post
        As far as I know RDNA3 has/had some issues too.

        The main difference vs RDNA2 is that it is dual issue architecture, but it's not always working. And it can't dual issue packed math - so in reduced precision workloads its peak performance is basically halved. Sadly, we don't have clpeak results for half precision here.
        It wouldn't be a problem if the cards were not nerfed vs RDNA2 - for example we have 4608 cores in 6800XT vs 3840 cores in 7800XT.
        rdna3 can only dual issue 32 bit instructions with at most 2 operands limiting the instructions that dual issue can be used on, especially fma cant be dual issued in fp32, so you have to use wmma to get decent gemm performance (which clpeak cant).

        Originally posted by sobrus View Post
        As a sidenote, I wonder what they are doing.
        - they have CDNA which is strong at compute and with AI (WMMA instructions) but not for consumers
        - and RDNA which doesn't have WMMA and only does FP16 (2x faster than FP32)
        - and RDNA2 which doesn't have WMMA but is strong at general compute and rapid packed math down to INT4 (8x faster than FP32)
        - and RDNA3 which is poor at packed math but has WMMA for AI.
        - and RDNA4 which again will be something different.
        cdna dose not have wmma it has mfma a different and much more capable set of instructions backed by significantly more dedicated hardware.
        rdna2 cant really do int4 or int8 faster its just a couple of dprod instructions that can do that.
        rdna3 can do the same dual fp16 instructions as rdna2 so it isent worse in any way.
        Last edited by DiamondAngle; 12 December 2024, 05:30 PM.

        Comment

        • tracker1
          Junior Member
          • Nov 2019
          • 13

          #14
          I'm hoping that Intel can work out their issues and that they are at least breaking even on the B580. The GPU surface is massive and this must be pretty expensive to produce. They've priced it well for consumers, but looking at the Phoronix testing, this was obviously meant to be more of a high end card.

          I'm also hoping that the drivers stabilize quickly.

          Not to mention, if AMD prices the 8000 series well, it would be nice to see the needle move against NVidia.

          Comment

          • sophisticles
            Senior Member
            • Dec 2015
            • 2591

            #15
            Originally posted by L_A_G View Post
            Having been familiar with LTT from the early days of Youtube I really wouldn't place much stock in a guy who barely knows his ass from his elbow and is only successful because he started early. Especially when the guy's promoted more scientifically illiterate crap than I can count (stuff like wireless power products with sub 5% efficiency).
            I assume you guys are talking about Linus' Tech Tips.

            He has claimed that he was offered 100 million dollars for LTT and turned it down.

            Comment

            • smitty3268
              Senior Member
              • Oct 2008
              • 6955

              #16
              Originally posted by sophisticles View Post

              I assume you guys are talking about Linus' Tech Tips.

              He has claimed that he was offered 100 million dollars for LTT and turned it down.
              Yeah and Elon had an offer to take Tesla private until suddenly he didn't.

              Linus has zero tech credibility for me, but I do respect his hustle. The guy knows how to make money.

              Comment

              • jaxa
                Senior Member
                • Jul 2020
                • 352

                #17
                Originally posted by sobrus View Post
                As a sidenote, I wonder what they are doing.
                - they have CDNA which is strong at compute and with AI (WMMA instructions) but not for consumers
                - and RDNA which doesn't have WMMA and only does FP16 (2x faster than FP32)
                - and RDNA2 which doesn't have WMMA but is strong at general compute and rapid packed math down to INT4 (8x faster than FP32)
                - and RDNA3 which is poor at packed math but has WMMA for AI.
                - and RDNA4 which again will be something different.

                I wonder how they could do FSR using AI if every single their GPU generation has different compute properties?
                Apparently, what they are doing is merging RDNA and CDNA into a new unified architecture called UDNA, and while it hasn't been officially announced, UDNA1 may replace what would have been called RDNA5.

                Comment

                • sophisticles
                  Senior Member
                  • Dec 2015
                  • 2591

                  #18
                  Originally posted by smitty3268 View Post
                  Yeah and Elon had an offer to take Tesla private until suddenly he didn't.
                  Actually what Elon said was that he was taking Tesla private at $420 a share. as a joke for his pot loving girlfriend.

                  The SEC was not amused and fined him 10 million bucks for his little joke.

                  Comment

                  • DOSftw
                    Junior Member
                    • Apr 2019
                    • 3

                    #19
                    Originally posted by L_A_G View Post

                    Having been familiar with LTT from the early days of Youtube I really wouldn't place much stock in a guy who barely knows his ass from his elbow and is only successful because he started early. Especially when the guy's promoted more scientifically illiterate crap than I can count (stuff like wireless power products with sub 5% efficiency).
                    I agree, but at least they understand this and basically openenly declare themselves as a "entertainment" content. People watch them not for a scientific benchmarks or analysis but rafther to be amused in some "IT/HW set".

                    I couldn't say the same about the Coreteks though. The dude thinks he has some seriuos intelectual insights/speculations but spreads mostly garbage misleading takes and fantasy level wishful thinking.

                    Comment

                    • Eudyptula
                      Phoronix Member
                      • Nov 2023
                      • 67

                      #20

                      Originally posted by jaxa View Post

                      Apparently, what they are doing is merging RDNA and CDNA into a new unified architecture called UDNA, and while it hasn't been officially announced, UDNA1 may replace what would have been called RDNA5.

                      https://www.tomshardware.com/pc-comp...cuda-ecosystem
                      This is great news. I just wished something like this happened a while ago, not a while into the future. Still, it's great to see that they are making good decisions for the future.



                      Originally posted by sophisticles View Post

                      Actually what Elon said was that he was taking Tesla private at $420 a share. as a joke for his pot loving girlfriend.

                      The SEC was not amused and fined him 10 million bucks for his little joke.
                      There was a "funding secured" comment as well. Along with some other comments. He likes to make comments. Many of which come back to bite him (and then shoot him in the foot (but then some other comments and investments come back and patch up his foot and help fight off the biting comments (this cycle repeats indefinitely in nth elon units (how many nested parantheses are allowed, exactly?)))).
                      Last edited by Eudyptula; 13 December 2024, 06:18 AM.

                      Comment

                      Working...
                      X