Early Linux 6.12 Kernel Benchmarks Showing Some Nice Gains On AMD Zen 5

[phoronix]

Phoronix: Early Linux 6.12 Kernel Benchmarks Showing Some Nice Gains On AMD Zen 5

With the Linux 6.12 merge window wrapping up this weekend and the bulk of the new feature merges now in the tree, I've begun running some Linux 6.12 benchmarks. Here is an initial look at Linux 6.10 vs. 6.11 vs. 6.12 Git on an AMD Ryzen 9 9950X desktop...

https://www.phoronix.com/news/Linux-6.12-Early-AMD-Zen-5

[Errinwright]

Not shabby whatsoever

[mlau]

zen5 is quite impressive, my test program (parse a 256MB csv into multiple linked lists the naive way) gets throughput of 5,5insn/clock:

Performance counter stats for './ijiximport-pgo ijix-w2':

           182,85 msec task-clock:u                     #    0,846 CPUs utilized              
                0      context-switches:u               #    0,000 /sec                      
                0      cpu-migrations:u                 #    0,000 /sec                      
           80.335      page-faults:u                    #  439,347 K/sec                      
      613.751.179      cycles:u                         #    3,357 GHz                        
       53.635.470      stalled-cycles-frontend:u        #    8,74% frontend cycles idle      
    3.388.648.732      instructions:u                   #    5,52  insn per cycle            <=======
                                                 #    0,02  stalled cycles per insn    
      846.197.599      branches:u                       #    4,628 G/sec                      
        1.696.638      branch-misses:u                  #    0,20% of all branches

on zen4, maximum is 4,5 insn/clock.

[NeoMorpheus]

I would love to see these tests between the Zen 4 and Zen5 equivalent CPUs, to see the overall uplift.

[jonkoops]

I don't understand the hate these CPUs get from the press, gaming performance is just fine, and these things are productivity and efficiency monsters. There should not be a need to upgrade every year anyways.

[entropicdrifter]

I don't understand the hate these CPUs get from the press, gaming performance is just fine, and these things are productivity and efficiency monsters. There should not be a need to upgrade every year anyways.

Peoples' expectations are colored by the explosive generational leaps of yesteryear. This generation's backlash reminds me of the backlash against IvyBridge being barely better than SandyBridge (and worse if you were an overclocker). IMO it's just noise

[69Y69]

I don't understand the hate these CPUs get from the press, gaming performance is just fine, and these things are productivity and efficiency monsters. There should not be a need to upgrade every year anyways.

Zen 5 was expected (and advertised) as the first breakout from Zen 1 lineage. This set the expectations baseline above previous gen perf gains.

Yet, the overall gain doesn't seem to match Zen 3 or Zen 4 gains.

[geerge]

Zen 5 was expected (and advertised) as the first breakout from Zen 1 lineage. This set the expectations baseline above previous gen perf gains.

Yet, the overall gain doesn't seem to match Zen 3 or Zen 4 gains.

Well, it is the first major redesign. However it's normally the gen after with refinements that usually is where stuff hits its stride. Zen was just ok, great relative to where AMD was before sure, but it took until Zen2 for AMD to really start cooking. As usual AMD's marketing is more a burden than a benefit.

[coder]

I don't understand the hate these CPUs get from the press, gaming performance is just fine, and these things are productivity and efficiency monsters.

There were some early problems with scheduling bugs, due to the Windows drivers leaving sticky state between CPU swaps, and some kind of quirk that impaired branch prediction when not run as the admin user. These meant you only got a decent uplift if you did a fresh Windows reinstall between testing different AM5 CPUs on the same system and ran your tests as the admin user.

The other complaint people had about them is the perf/$ value, as compared with the street prices of Ryzen 7000 CPUs. That's a little unfair, IMO. Perf/$ always gets a little worse, towards the top end of the curve, and it's also normal for new products to have a bit of a premium. Furthermore, when these are outpacing Intel in some key areas, there's going to be a premium for that.

Finally, if you're not comparing against the 9900X or 9950X, Raptor Lake's multi-threaded performance looks better, due to their hybrid architecture. This further hurts the value proposition of the 9700X, for instance.

We shouldn't lose sight of the fact that Zen 5 currently has the best single-threaded performance. The 12-core and 16-core models are also top performers on multi-threaded. I think that warrants a bit of a price premium.

I think Arrow Lake is going to be in a similar situation vs. Raptor Lake, where it's not a huge increase on either single-threaded or multi-threaded performance. So, we'll see how similar the press' reaction to it is.

There should not be a need to upgrade every year anyways.

I think people got spoiled by how big the gains were in Alder Lake and Raptor Lake. Ryzen 7000 also delivered pretty good gains, with higher TDPs and DDR5 doing a lot to alleviate bottlenecks the Ryzen 5950X faced on multithreaded performance.

Something else to consider is that the rumor mill ran for a long time about how Zen 5 was a big redesign and I think that created an expectation that the gains should be even bigger than some of the prior generations. If the expectations had been set low, then maybe the overall impression would be better.

[coder]

Yet, the overall gain doesn't seem to match Zen 3 or Zen 4 gains.

IPC-wise, it actually delivered bigger gains. That's where you look for them, on such a big microarchitectural reset. It's not surprising that clockspeeds didn't move much, given they went from N5 -> N4P (which is an N5-derivative and not a fundamentally new node).

It sounds to me like there's plenty of room for further optimizations in Zen 6. Both from further node shrinks and just having an opportunity to do lots of tuning in the new microarchitecture. That said, AMD recently proclaimed themselves to be a Datacenter-first company. We should therefore continue to see them focus on scalability and efficiency over and above high clock speeds.

https://seekingalpha.com/article/4720050-advanced-micro-devices-inc-amd-goldman-sachs-communacopia-and-technology-conference

Zen cores have never been as big and complex as corresponding Intel cores. This really helped AMD scale core counts aggressively, in their EPYC lineup, but it has put them at a slight disadvantage on IPC vs. Intel. They also don't have a separate desktop core, like Intel does, and therefore it's not as frequency-optimized.

On the flip side, the 3D V-cache is something AMD did just for servers. They only created the consumer X3D products, when one of their employees decided to see how much it benefited gaming performance.