Announcement

Collapse
No announcement yet.

Intel Updates Alder Lake Tuning For GCC, Reaffirms No Official AVX-512

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

  • #21
    Originally posted by avem View Post

    You can perfectly use AVX-512 after disabling E-cores and AVX-512 tasks run circles around AMD CPUs.
    Perfectly? Even though they're not validated?
    gentoofu
    Junior Member
    Last edited by gentoofu; 10 November 2021, 01:46 PM.

    Comment


    • #22
      Originally posted by drakonas777 View Post

      OK, so it's 17-30. Depends on the workload - that's why I gave a rough range and not an exact number. The point is it's non-trivial. Also, I was writing about MT efficiency, not low threaded tasks. AL efficiency across mixed workloads at certain power limits is indeed better, but PL2 of 12900K is just not practical in MT. What's the point to pay premium for i9 K and then limit the power. Choosing non-K potential future i9 or at least current 12700K makes much more sense.

      My main criticism is that 12900K PL2 is insane and kind of misleading. It's basically "benchmarking mode". Also, PC tech community seems not to care much and does not call it out generally speaking, which means raising power limits to insane levels for the mainstream parts will become a norm in the future. It all kind of started with Ampere. Now AL i9. I'm not a fan of this tendency for a lot of reasons - but it's offtopic, yes.
      From my calculation it is -6.7% performance from 241W to 160W (25590/27430 = 0.933). So not sure how you are getting to 17%.

      Other than that, I also don't like these SKU tuning for maximum performance. Setting it to 125-150W would be much more sensible in terms of efficiency without a tremendous performance loss. The power increase trend of CPUs and GPUs should really stop, as it completely goes against all climate and sustainable energy goals! This is an area where "competition" has actually made the situation worse.





      Comment


      • #23
        Originally posted by Michael_S View Post
        I remember a few years back there were people in these forums and a few other tech discussion sites loudly proclaiming that Intel was crushing AMD because of Intel processor AVX-512 support. They said it didn't make sense for anyone to get AMD parts because they didn't have it, and AVX-512 was - in their minds - critical for lots of software. So were they fanboys, or paid shills? (Probably just fanboys, but it was mind-boggling to see people insist this feature was so important when it was brand new and nothing used it.)
        Broadly speaking I was one of them (but I never said Intel was gonna crush AMD), because longer vector length delivers higher performance in linear algebra (you know what I mean if you have used scientific computing software).

        But no body (at least not me) expected Intel failed to deliver the anticipated performance. There is even no guarantee that AVX-512 units will work at BASE frequency, according to the official document. What a shame.
        zxy_thf
        Senior Member
        Last edited by zxy_thf; 10 November 2021, 02:55 PM. Reason: typo

        Comment


        • #24
          Originally posted by gentoofu View Post

          Perfectly? Even though they're not validated?
          Lots of people claim to run ECC RAM in RyZen systems where AMD doesn't actively turn off ECC but it's not actively supported either. Half the time the ECC isn't even enabled, but it doesn't stop them.

          Comment


          • #25
            Originally posted by gentoofu View Post

            Perfectly? Even though they're not validated?
            Have you actually read the entire tweet? There's no validation program for AVX-512. Meanwhile do you have actual proofs that an AVX-512 implementation in ADL computes incorrectly? No? Then WTF are you talking about?

            WTF is wrong with people? Can you start discussing GCC patches for ADL and stop shitting on Intel and its new CPU lineup? Don't you have better things to do in life?

            Comment


            • #26
              Originally posted by cchi View Post

              From my calculation it is -6.7% performance from 241W to 160W (25590/27430 = 0.933). So not sure how you are getting to 17%.

              Other than that, I also don't like these SKU tuning for maximum performance. Setting it to 125-150W would be much more sensible in terms of efficiency without a tremendous performance loss. The power increase trend of CPUs and GPUs should really stop, as it completely goes against all climate and sustainable energy goals! This is an area where "competition" has actually made the situation worse.
              It's 100% workload dependent. Low threaded / background tasks might not be affected if your limit its TDP to even 65W.

              The 241W figure was chosen to beat 5900X and rival 5950X. Was it a good idea? I don't know but AMD fans are all over it as if someone makes them buy this CPU. The owners of this CPU are quite content with it. It's a lot easier to cool than Ryzen 5000 CPUs because their chiplets are tiny and very hard to take off heat from.
              Last edited by avem; 10 November 2021, 06:04 PM.

              Comment


              • #27
                Originally posted by Michael_S View Post

                The news post is also about Intel reaffirming that their new flagship won't have supported AVX-512, it's right in the headline - which you conveniently overlooked.

                Yes, I was joking around in my post. But my point is that a technology Intel heralded as hugely significant at introduction has been deemed by Intel as not significant. That change is, if you'll pardon the repetition, significant.

                Edit: and mocking fanboyism is not itself fanboyism. The room I'm in has more and faster Intel parts than AMD.
                Show me any official marketing materials prior to the ADL release where Intel promised or advertised AVX-512. Speaking frankly I don't understand what the heck you're talking about. Some people had some expectations based on some rumors and now those expectations aren't even that far off since you still can technically use AVX-512.

                What's the fuss about exactly? The people who actually need AVX-512 can go ahead and disable e-cores permanently - this looks like a win-win situation. You lose some, you gain some and Intel sells you something which is not even there but still works.

                Comment


                • #28
                  Originally posted by chuckula View Post

                  Lots of people claim to run ECC RAM in RyZen systems where AMD doesn't actively turn off ECC but it's not actively supported either. Half the time the ECC isn't even enabled, but it doesn't stop them.
                  ECC not being validated (or not working at all) is actually a huge deal because e.g. in datacenters there are literally terabytes of RAM and bit flips due to cosmic radiation are a common occurrence ECC deals with.

                  I've had bit flips on my system with 64GB of RAM which was quite a surprise for me. I've always thought of them as something extremely rare and not affecting mere mortals.

                  More info:

                  https://newbedev.com/cosmic-rays-wha...fect-a-program
                  https://blogs.oracle.com/linux/post/...he-cosmic-rays
                  https://www.johndcook.com/blog/2019/...flipping-bits/
                  Last edited by avem; 10 November 2021, 06:14 PM.

                  Comment


                  • #29
                    Good lord, I think birdie wrote about half the comments under this article.

                    I'll never understand why some people get so invested in a company.

                    Comment


                    • #30
                      Originally posted by avem View Post
                      It's power hungry only for massively parallel computational tasks.
                      This is not accurate. I see you quoted from the same article, but maybe you missed the core-scaling chart:



                      Originally posted by avem View Post
                      In every day common tasks it's the most power efficient high performance x86-64 CPU out there.
                      You really can't call burning 78 W (+71 W above idle) for a single-threaded task particularly efficient, much less "the most power-efficient". I mean, maybe you're joking but I'm not detecting a hint of sarcasm.

                      That's about the same as the peak all-core load on a Ryzen 5600X uses, and you'd better believe that's going to deliver more aggregate performance.

                      Comment

                      Working...
                      X