No announcement yet.

AMD Ryzen 9 7900X / Ryzen 9 7950X Benchmarks Show Impressive Zen 4 Linux Performance

  • Filter
  • Time
  • Show
Clear All
new posts

  • #51
    AMD have OCed Ryzen 7000 CPUs too much. The CPUs would have shone if AMD had slashed their TDP by just 20-30%.

    Last edited by birdie; 26 September 2022, 06:41 PM.


    • #52
      Originally posted by ojab View Post
      Can I haz 7950X benchmark in ECO mode & TDP limiting? Anandtech says that
      170W TDP Base = 105W with ECO Mode enabled
      105W TDP Base = 65W with ECO Mode enabled​
      so there are 3 possible combinations except default and probably 170W+ECO is the same as plain 105W limit
      These results are disappointing since they confirm what I already suspected, that the 7950X is not more power-efficient than the 5950X... however, ECO mode could potentially redeem the 7000 series. Would be nice to see it retested with ECO mode.

      Edit: actually, no need, I got the number I was looking for. When lowering the TDP from 170W to 65W, multi-threaded performance drops by about 18%. That is crazy good efficiency.
      Last edited by david-nk; 26 September 2022, 07:44 PM.


      • #53
        Does anyone around here know if Ryzen 7000 has Pluton in it or not? Anandtech is the only site I've found that's even mentioned the possibility of it being in there and I figured Linux diehards would care more about this sort of thing. AMD didn't mention Pluton at all in their desktop Zen 4 press release (whereas they were more than happy to use it as a selling point when they announced Ryzen 6000 + Ryzen 7020), I didn't see it show up in any of the slides they published today and I remember Robert Hallock saying he didn't know if Pluton was in Zen 4 when someone asked him about it a couple of months ago. I'm assuming Pluton is in these CPUs, but these things are coming out tomorrow and no one's bothered to confirm or deny it yet, not even AMD themselves.

        Most of the coverage on these processors seems to be exclusively focused on "lol these things are space heaters!!", which is a valid concern, but it's made it even harder to try and dig up any kind of Pluton-related info at all and I was hoping at least a few journalists would be paying attention to it. D:


        • #54
          Originally posted by atomsymbol
          Increasing IPC is not so hard if the CPU will run at a lower frequency (such as: 1 GHz). It is hard if the goal is to increase IPC as well as increase frequency.
          That's not true, independent of ISA. x86 has something of a bottleneck in the decoder, which means you need to clock it higher to get good throughput from it. Conceivably, you might double-clock it or run it asynchronously from the back end, but that probably requires adding another pipeline stage.

          In the backend, you can widen things a bit, but you're still constrained by ILP (instruction-level parallelism) and that's constrained by the number of ISA registers (among other things).

          Taken together, both points mean x86 tends to favor higher-clocked, narrower cores than something like ARM.

          Originally posted by atomsymbol
          According to AMD's hints about their next CPU: Zen 5 (Ryzen 8000?) underwent a "complete" redesign of the core which increases the probability that Zen 5 will have ~20% higher IPC than today's Ryzen 7000.
          Well, if you double the width of Zen's vector pipelines, then your peak theoretical FP IPC could double. So, if we're counting that, then there's room left. If we're talking about integer or scalar IPC, that's much harder.

          Originally posted by atomsymbol
          Intel Raptor Lake CPUs are expected to increase single-threaded IPC by ~10%.
          Maybe single-threaded performance, but not IPC. It's the same microarchitecture. You don't gain that much IPC with bigger caches and faster DDR5.


          • #55
            Originally posted by birdie View Post
            AMD have OCed Ryzen 7000 CPUs too much. The CPUs would have shone if AMD had slashed their TDP by just 20-30%.
            If you care about perf/W, then you're going to run it in Eco mode, with a lower TDP limit.

            Outside of that, of course they're going to burn more power for a little more performance, because that's the game Intel is playing.


            • #56
              I would rather see a perf/watt bios setting than a baked in perf/maxT algorithm. Good that you can select maxTDP though.


              • #57
                Originally posted by Mahboi View Post
                I'm very confused. How do we go from a TSMC 7nm drawing between 65W-150W to a TSMC 5nm drawing 105W-250W, two years of work from AMD on IPC, a whopping 1Ghz improvement across all processors, and we end up with a broad 15-20% increase in most benchmarks? Yes some are stellar (compilation and some AVX-512 related things), but if my poor brain can math for once in my life:

                7nm to 5nm: expected roughly 30% performance increase or lower consumption
                150W top to 250W top: 66% wattage increase
                2 years of work and a promised ~8% IPC increase
                1Ghz on top of a broadly 4Ghz lineup: 20-25% increase

                So I count 8% IPC + 20% speed + 30% lower consumption + 66% wattage increase. Even if you know that more watts != more speed, I'm still baffled here. How are we seeing such a massive power draw and such a "good" growth of speed? Since some benchmarks are incredibly faster, I'm thinking it could be a microcode question, that AM5 and Zen4 are entirely new and that AMD will have months of optimisation in the coming months/years to fully exploit them, but otherwise, I really really do not understand how these numbers add up.

                In any case, my little 5600x is doing amazing in perf/watts, so I'm very very happy I bought it.
                It's because ZEN4 is ZEN3 derivative. AMD even said this during official unveil event. In order to get significant improvement uarch must be fundamentally redesigned. ZEN5 should be the one which brings far more bigger bump.


                • #58
                  Originally posted by atomsymbol
                  Theoretical (frequency-independent) ILP is constrained by exactly one limit: temporal dependencies among bits of data.
                  And the compiler is limited in its ability to express those dependencies by the number of ISA registers. Sure, you could have a CPU which identifies spills, but there are practical limits on what can be efficiently implemented in hardware. This stuff not only takes die area, but also energy and time.

                  Originally posted by atomsymbol
                  ILP is not about bigger caches, nor about faster DDR5.
                  Sure it is. Reducing memory stalls increases the amount time a CPU core can spend getting useful work done.


                  • #59
                    Originally posted by jrch2k8 View Post
                    what i meant is sure, you can have a 1000w cpu that is more efficient than 65w cpu but that doesn't mean i'm comfortable with that power draw.
                    Also schmidtbag

                    cpupower frequency-set -u 5.2 GHz

                    The power is in your hands!


                    • #60
                      Originally posted by jrch2k8 View Post
                      95c and that power draw are a skip 7000 series for me. I think an used 5950x will be the sweet spot soon
                      According to PCWorld, you can enable eco-mode to massively reduce power draw while losing 0%-10% performance.