[mattbillenstein]

I came to realize by soldering a lot of fast ram on to the board of newer laptops and phones, maybe it's not the instruction set that matters that much.

Modern Apple hardware has so much more memory bandwidth than the x86 systems they're being compared to - I'm not sure it's apples to apples.

[hajile]

A19 has WAY less bandwidth on its 64-bit bus than desktop chips with 128-bit busses . AMD’s strix halo is also slower despite a 256-bit bus.

Pushing this point further, x86 chips are also slower when the entire task fits in cache.

The real observation is how this isn’t some Apple black magic. All three of the big ARM core designers (Apple, ARM, and Qualcomm) are now beating x86 in raw performance and stomping them in performance per watt (and in performance per watt per area).

It’s not just apples deep pockets either. AMD spent more in R&D than ARM’s entire gross profit margin last I checked. Either AMD sucks or x86 has more technical roadblocks than some people like to believe.

[zamadatix]

Spot on about the memory, even some of the actual M* models don't have all that high of memory bandwidth and still kick ass just as well as the ones that do in this kind of benchmark.

I do feel like x86 has more technical roadblocks, but disagree the amount of investment is not the primary driving factor at this point. I haven't seen designs from ARM itself beat x86 on raw performance yet, and 100% of their funding goes towards this point. E.g. the X925 core certainly doesn't, nor does the top single core Android device on e.g. Geekbench come close to current iOS/PC device scores. They've announced some future shipping stuff like the C1 which is supposed to, but now we're talking marketing claims about upcoming 2026 CPUs vs Zen 5 from 2024. Perf/Watt wise absolutely of course, that ship sailed long ago. Z1/Z2 were admirable attempts in that regard, but still a day late and a dollar short to leading ARM designs.

The other factor to consider is scale-out CPUs with massive DC core counts tend to have mediocre single core performance, and that's what AMD really builds Zen for. Compare to Graviton in the DC and AMD is actually performing really well in both single/multi performance, perf/watt, and perf/dollar. It just doesn't scale down perfectly.

Apple/Qualcomm have certainly dumped more R&D into their cores being low core count beasts, and it shows vs any competition (ARM or x86). The news likes to talk about how many of the Nuvia developers came from working on Apple Silicon, but I think that is a bit oversold - I think it's mostly that those two development programs had a ton of investment targeting this specific use case as the main outcome.

[hajile]

x925 does according to GeekerWan. C1 Ultra is even faster. x86 GB6 results are from the Geekbench website. I searched for the fastest overall scores I could find in the first few pages of GB6 results to steelman as best as possible.

The long and the short is that x86 is WAY behind in every way. The chips are larger, hotter, and slower too. If the rumored A19 Pro in a $500 laptop happens, it's going to absolutely crush the wintel market.

The stuff about Graviton is missing a key element too. Look at the X3 scores below. They are around 30% slower than the x86 competitors. This is what Graviton 4 is using (Neoverse V2 is based on X3). Neoverse V3 was announced almost 2 years ago now and is based on X4 which is a pretty big jump. I'd expect Neoverse V4 in Feb 2026 to be based on either X925 or C1 Ultra. When these newer, faster cores hit the market, they will be beating x86 in cost (the cores are smaller) and power consumption if not peak performance too.

    CPU             Spec2017int Spec2017fp  GB6 1T   GB6 nT
    a19 Pro            12.11      17.37      4019    11,054
    A18 Pro            10.63      15.93      3605     9,376
    A17 Pro             9.39      12.92      2975     7,273[10]
    A16                 8.52      12.01      2679     7.026[11]
    9400 (x925)         8.73      13.67      3010     9,317
    9500 (C1 Ultra)    ~9.50     ~16.50      3709    10,716
    O1 (x925)           9.19      14.46      3125     9.671
    SD8Gen2 (X3)        6.52      10.49      2080     5,640[7]
    SD8+ (X2)           5.71       9.68      1656     3,716[6]
    SD8Gen3 (X4)        7.10      10.95      2366     7,552[9]
    8 Elite             8.91      14.18      3237    10,242
    8 Elite Gen 5      ~9.75     ~16.50      3946    12,546
    U7 258v             8.28      11.57      2864    11,263[3]
    HX370               8.02      12.81      3059    14,525[4]

[0] https://www.youtube.com/watch?v=3PFhlQH4A2M

[1] https://www.youtube.com/watch?v=YJaHi-gZESo

[2] https://www.youtube.com/watch?v=cB510ZeFe8w

[3] https://browser.geekbench.com/v6/cpu/14113967

[4] https://browser.geekbench.com/v6/cpu/14136651

[5] https://www.notebookcheck.net/AMD-Zen-5-Strix-Point-CPU-anal...

[6] https://browser.geekbench.com/v6/cpu/14141425

[7] https://browser.geekbench.com/v6/cpu/14173239

[8] https://web.archive.org/web/20240308050136/https://www.anand...

[9] https://browser.geekbench.com/v6/cpu/14169554

[10] https://browser.geekbench.com/v6/cpu/14172509

[11] https://browser.geekbench.com/v6/cpu/14172532

[lordnacho]

I talked to a guy who'd worked at Apple on the chips. He more or less said the same thing, it's the memory that's all the difference.

This makes a lot of sense. If the calculations are fast, they need to be fed quickly. You don't want to spend a bunch of time shuffling various caches.

[zamadatix]

Memory bandwidth/latency is helpful in certain scenarios, but it can be easily oversold in the performance portion of the story. E.g. the 9950X and 9950X3D are within less than 1/20th of a percentage point of each other in PassMark Single thread (feeding a single core is dead easy) but have a spread of ~6.4% (in favor of the 9950X3D) in the multi-thread (where the cache is starting to help on the one CCD). It could just as easily have been in the other direction or 10 times as much depending on what the benchmark was trying to do. For most day to day user workloads the performance difference from memory bandwidth/latency is the "nil to some" though.

Meanwhile the AI Max+ 395 has at least twice the bandwidth + same number of cores and comes to more like a ~15% loss on single and ~30% loss on multithread due to other "traditional" reasons for performance difference. I still like my 395 though, but more for the following reason.

The more practical advantage of soldered memory on mobile devices is the power/heat reductions, same with increasing the cache on e-cores to get something out of every possible cycle you power rather than try to increase the overall computation with more wattage (i.e. transistors or clocks). Better bandwidth/latency is a cool bonus though.

For a hard number the iPhone 17 Pro Max is supposed to be around 76 GB/s, yet my iPhone 17 Pro Max has a higher PassMark single core performance score than my 9800X3D with larger L3 cache and RAM operating at >100 GB/s. The iPhone does have a TSMC node advantage to consider as well, but I still think it just comes out ahead due to "better overall engineering".

[hamandcheese]

It's very possible I am misinterpreting, but the A19 seems to have less total memory bandwidth than, say, a 9800x (but not by much). And far less than the Max and Ultra chips that go into MacBooks.

So I think there's more to it than memory bandwidth.