Apple's M1 Max is 1.5x faster than M1 in supposed benchmark

2»

Comments

  • Reply 21 of 25
    mjtomlin said:
    mjtomlin said:
    mjtomlin said:
    These numbers are suspect. As I pointed out in the article about the GPU performance, GPU perf went up way more than 2x with 2x GPU cores, according to Apple's numbers (which, historically, they have not inflated). That strongly suggests that the cores are NOT M1 cores, but at least equivalent of the A15 cores.

    It would be absolutely astonishing for the single-core score not to go up at all. In fact I'd say it was totally impossible. If nothing else, the new memory controller, bigger SLC, and LPDDR5 memory will all increase the single-core score. Probably by not all that much, but claiming no change at all is a huge stretch. (And that assumes no clock bump over the M1.)

    The 150% bump for multicore might actually make sense with the 8-core version of the CPU (6P + 2E cores). It too seems way too low for the 10-core version of the M1 Pro.

    We'll know soon enough though.

    Actually Apple claimed 2x performance with 16 GPU cores, and 4x performance with 32 cores. The M1 has 8 cores. GPU performance scales linearly. CPU performance does not. Apple said the new 10-core CPU is 70% more performant then the M1. This GB score is just outside that, but that’s completely possible as these benchmark scores are usually all over the place.

    And that single core number is absolutely possible if Apple is using the same cores as the M1, which they most likely are, or they would’ve called it M2 to signify different core generation as they do with the A-series. It was rumored that these SoCs were ready to go mid-year, but there were delays with the displays, so the release was pushed back.
    What you wrote makes sense... but it's wrong.

    If you look at the GPU performance chart Apple showed during the presentation, it compared the M1 Pro and Max GPU *not* to the M1, but rather to the 5600M in the most recent intel 16". That 5600M is markedly faster than the M1 GPU. Apple numbers were improvements of 250% and 400% respectively. You can conservatively estimate the 5600M as 50% faster than the M1 8-core - that's *very* conservative, as geekbench has it as about 200% in metal, for example. That would make the M1 Pro GPU 3.75x the speed of the M1, and the Max would be 6x faster. The most obvious way to explain this would be better memory bandwidth and SLC, but even so I doubt all that performance is coming just from a better memory subsystem. Go look at the multicore improvement in the A15 vs. A14... it's suggestive, at least. Of A15 cores, or maybe the uncore, but... something.

    Of course, that also raises the question: GPU perf *should* increase linearly (as memory bandwidth is doubling too), but according to Apple it only goes up 60%. That is a big mystery, and I really want to know more about that.

    Lastly, as to CPU perf: even if it were the same core as in the M1 (which I doubt), performance would *have* to go up unless they actually lowered the clock to *slower* than the M1, as I explained in the quote above.

    Actually I was not wrong. During the keynote they said Pro GPU was “2x faster GPU performance than M1”, Max was “4x faster GPU performance than M1”. It had nothing to do with any chart, it’s what they said. Watch the keynote. This would indicate they are using the same GPU cores as the M1. Also, the M1 GPU has a TDP of 10W, the Max has a TDP of 40W… 4x the power draw. The newer GPUs in the A15 are more performant and more efficient - these are not those. So, it would stand to reason that they’re also using the CPU cores as well. And the 30W TDP is inline with those cores. Furthermore, they specifically said these SoC’s are based on the M1.

    [graphics removed]
    OK, good answer. And if the numbers you're quoting turn out to be correct, then obviously you're right about the cores. However, my numbers also come from graphics they showed during the presentation. So... Apple made two statements that strongly conflict with each other. Even more in conflict than the raw performance claims are the scaling claims - the numbers you quoted are as expected (linear for the GPU) whereas the ones I quoted were not. Guess we'll see which ones are right soon enough.

    I don't think the TDP numbers are conclusive, for the same reason. And there's still the major point that the memory subsystem should improve CPU single-core scores (geekbench 5, AFAIK, doesn't live entirely in cahce). But again, we'll see soon enough.

    As far as CPU benchmarking is concerned, I strongly think load/store latency is factored out otherwise you’re not really testing just the CPU. Data “bottlenecks” should not be a concern when you’re testing raw performance… Real world performance, yes, but not raw performance.

    If you have slow RAM or low cache, you can’t fault CPU cores for that. All you want to time is the core’s ability to compute.

    And yet, pretty much all benchmarks do test that, because it really is a significant part of the CPU. It's not like anyone cares (aside from academically) how well the branch predictor works. They care about results. That applies especially to the M1s, since a big part of their performance really is all about bandwidth and latency (the memory interface and the SLC) and reducing memory access (L2C and SLC). In fact, one of the things people look for in modern benchmarks is making sure they aren't small enough to fit in cache, because that warps the results too much to be useful for general performance profiling. (Of course, when you're interested in one very specific application, then you usually hope it *does* fit in cache, but that's a different story.)

    In the real world, the memory subsystem is key. Realizing the implications of this was crucial to Apple's CPU/SoC successes in recent years.
    edited October 2021
  • Reply 22 of 25
    canukstormcanukstorm Posts: 2,560member
    crowley said:
    red oak said:
    Have to be honest, this benchmark seems low if it is true.   With everything they launched today (impressive),  I thought we would see single core at 2,000 and multi at 14,000+ 

    We'll know the real answer by end of this week 
    Why would the single core score go up?  It's the same M1 cores, just more of them.  You'll need to wait for M2 for the single core needle to move substantially.
    And for multi they went from 6 high performance cores to 8, a 33% increase in the number of cores; so a 50% improvement seems better than you might otherwise expect.

    The real improvements are in the memory capacity and bandwidth, the GPU, and possibly the other processing units for video and neural.
    Apple's M1 has 4 high-performance cores (not 6) and 4 efficiency cores.
    watto_cobra
  • Reply 23 of 25
    mjtomlinmjtomlin Posts: 2,504member
    A lot more scores have shown up... ranging from 11,000 to almost 13,000. The highest so far is 12,693.

    This puts it on par with Core i9-7940X CPU, with 14 CPU cores that boost to 4.3GHz and has a TDP of 165W.
    Compared to 3.2GHz and a TDP of 40W for 10 CPU cores on M1 Pro/Max.
    edited October 2021 fastasleepwatto_cobra
  • Reply 24 of 25
    thttht Posts: 4,201member
    mjtomlin said:
    A lot more scores have shown up... ranging from 11,000 to almost 13,000. The highest so far is 12,693.

    This puts it on par with Core i9-7940X CPU, with 14 CPU cores that boost to 4.3GHz and has a TDP of 165W.
    Compared to 3.2GHz and a TDP of 40W for 10 CPU cores on M1 Pro/Max.
    You'd think the reviewers would run GB without having the scores logged into the database, to keep something for their reviews. They don't talk about performance benchmarks anyways, so no big loss I suppose.

    Anyways, for some perspective, a ~12000 score is between a Mac Pro 12c Xeon W-3235 and an iMac Pro 18c Xeon W-2191B. This is macOS using its Xcode toolchain.

    For Windows toolchains, or it could be Linux here too, ~12000 is comparable to 12c Ryzen 3900X and a bunch of rebranded Xeons as Core X processors. The 8c Ryzen 5800X is ~10000, 12c Ryzen 5900X is 14000. Intel Core i9-11900KF is ~10000. All desktop, workstation or server processors.

    For laptop processors, Ryzen 5900HK is 7600. The Core i7-11800H is 8000. It's all moving targets since it is race to the next node or next architecture.

    So, really good showing on the CPU front, especially if it is a true-blue 30 to 40 Watts, and not with an initial 60 to 70 Watt turbo for a minute or two.
    fastasleepwatto_cobra
  • Reply 25 of 25
    mjtomlinmjtomlin Posts: 2,504member
    tht said:
    mjtomlin said:
    A lot more scores have shown up... ranging from 11,000 to almost 13,000. The highest so far is 12,693.

    This puts it on par with Core i9-7940X CPU, with 14 CPU cores that boost to 4.3GHz and has a TDP of 165W.
    Compared to 3.2GHz and a TDP of 40W for 10 CPU cores on M1 Pro/Max.
    You'd think the reviewers would run GB without having the scores logged into the database, to keep something for their reviews. They don't talk about performance benchmarks anyways, so no big loss I suppose.

    Anyways, for some perspective, a ~12000 score is between a Mac Pro 12c Xeon W-3235 and an iMac Pro 18c Xeon W-2191B. This is macOS using its Xcode toolchain.

    For Windows toolchains, or it could be Linux here too, ~12000 is comparable to 12c Ryzen 3900X and a bunch of rebranded Xeons as Core X processors. The 8c Ryzen 5800X is ~10000, 12c Ryzen 5900X is 14000. Intel Core i9-11900KF is ~10000. All desktop, workstation or server processors.

    For laptop processors, Ryzen 5900HK is 7600. The Core i7-11800H is 8000. It's all moving targets since it is race to the next node or next architecture.

    So, really good showing on the CPU front, especially if it is a true-blue 30 to 40 Watts, and not with an initial 60 to 70 Watt turbo for a minute or two.

    Never seen my Mac mini go past 3.2GHz even under heavy load... there is no "boost" mode. And I think my math was off, it should only be ~30W for the CPU; ~3W each p-core, and ~0.75W for each e-core.
    fastasleepwatto_cobra
Sign In or Register to comment.