The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster

The source is Anandtech running SPEC on them, already given above.

SPEC is not a microbenchmark. It's been the industry standard for comparing completely dissimilar CPU architectures for decades.

I though you were talking specifcally about the low-power cores (and I couldn't find a site specifically comparing those, but even you could find those, it would be a weird niche (micro) benchmark, because it's not clear if those cores are used like that.

In any case, the link you shared (https://www.anandtech.com/show/16192/the-iphone-12-review/2) backs up what I said; the overall SoC perf of the 865 is quite close to that of the A13 (which is it's same-gen competitor). The specfp score of the 865 is 74% of the A13; the specint 64% (the 865+ is 3 percentage points better here). And the joules consumed by the A13 is significantly higher than the 865 for both benchmarks (the 865+ uses more energy than the 865, but less than the A13).

In fact, the geomean of perf and power between A13 thunder and lightning cores is pretty close to on the money for the 865, for both power and perf!

All in all, the 865 looks pretty competitive in the benchmark you linked; pretty much smack between the high-power and low-power cores of the A13: more efficient than it's high-power cores, but slow, and faster than the A13's low-power cores, but less efficient. Unfortunately, the page doesn't include the 865's low-power cores in the benchmarks, but from the point of view of disproving the apple-domination story it doesn't really matter: clearly competitors can build cores that are close to the cutting edge of apples power/perf tradeoff; they just choose slower, more efficient designs.

Where are you getting a 4x improvement in performance?

Ah, right I see what you're saying. However, note that these other SoC's don't rely on A55 (not alone - this is what I meant by microbenchmark; as it's comparing a small bit of the chips in isolation in a way you'd never normally use them), and that the A14 should be compared to the 888 if you want to tease out the design bit (apple's) from the process bit (TSMC's). Early 888 benchmarks look good, but it's more reliable to compare the 865 to the A13 since those numbers are clearly solid by now (no unreleased SoC shenanigans). And the 865's high power core is clearly competitive with the A13; it's slower (around 3/4 speed), but more efficient, not less than the A13 - again, the link you posted shows that.

So even if the low power cores in the A14 are impressive; clearly other competitors are capable of getting results that are competitive with apple on the same node - at least for the high power cores.

I strongly suspect the low power cores just aren't a priority for qualcomm; but in any case when in comes to servers (what this all started with), clearly the high-power cores are the ones that matter, and it's equally clearly not the case that apple's lead looks unassailable. It's been shrinking year by year, not growing, and the difference isn't as large as it's made out to be.

Looking at this data I get the impression that Apple has more resources to throw at the problem, and that some parts of their solution are simply better - but the difference is quite small, and they use more transistors to get there, and need more power planes (thus cost) to get the greater efficiency. The big difference is simply how much head start they get at TSMC, which is a question of money, not some secret design sauce.

And again this all started on a thread about server CPUs and memories of how slow chromebooks were (i.e. not talking of efficiency cores in isolation, but SoC perf overall). And for that I think data speaks for itself: Apple has a significant lead - but a small one, that's been shrinking over the years. There is no evidence they're in a class of their own at the same process node - other chips are 3/4 as fast and more efficient, and that seems like a reasonable tradeoff. Nor incidentally is this just a 2 horse race; much as samsungs chips are derided, they're not that much worse, and likely much of that is due to the inferior process node. Huawei too seemed competitive pre-trade-war, sometimes beating qualcomm. Given that at the same process node there are 3 different competitor that come so close, it just does look like Apple's design is really all that unique. It's quite conceivable a different ARM competitor might catch up in a few years, given the current trends. Then again - only if they get a slot at TSMC, since there are rumors that Apple has already bought much of the 3nm production, and as intel showed in the past - decent design with a process node advantage is a winning combination. But Apple needs that process node advantage to keep a large lead.