Apple gets a lot of flack for the high margins on their products, but those high margins allow them to design and create high-cost SoCs. They also essentially have a business guarantee that all of the chips they fabricate will get purchased, at a know price and ROI.

So, on top of the vertical integration of software/hardware advantage, they have diminished risk.

They’re also optimizing for UI/UX by emphasizing single more performance. That’s what affects how fast and responsive the interface is to user input.

After looking at this closer, what they are comparing is single threaded performance of the Skylake Xeon 8176. At 28 cores and 56 threads, that watts per thread or core value starts to look somewhat comparable, given Apple's low core count (compared to that desktop CPU) and that they might not run the Vortex (2x) and Tempest (4x) cores at the same time(?).

So, I think it's both a matter of comparing Apples to Oranges (ugh, I would use a less confusing analogy if one came to mind) in that the types and usage and number of cores are different, and also just fairly impressive that they can get that much single threaded performance out of the chip.

The craziest thing is how close they get at 3W. That's the part that's staggering. What could Apple do with, say, laptop thermals? Desktop thermals? The mind creeps closer to boggling.

It's not that staggering. Use the above example. It's comparable to one thread of the 56 thread Xeon 8176.

That server CPU has a TDP of 165W. That's 2.95W per thread. So pretty much entirely on par with Apple.

Well, except that hyperthreads are not the same as cores, so it's not quite equivalent. My understanding is that the two hyperthreads for a core share some execution resources, but are able to run in parallel for other resources. That means some workloads won't benefit much at all, others will benefit greatly, and the usual case is somewhere in-between.

Also, I think I've heard people here noting how hyperthreading causes some workloads to go slower (possibly from CPU heat throttling? Dunno) so some people disable it in the BIOS first thing.

So the truth is likely somewhere between 2.95W/thread and 5.90W/core (assuming 165W is even correct, there's comments here noting while AMD and ARM quote max, Intel quotes operating average...).

But that’s not a useful comparison, because you can’t run a Xeon at 2.95W. There is a floor wattage that the processor simply cannot go below.

User interface drawing/responsiveness depends most on single core performance.

IIRC one of the changes from the A10 to the A11 was that the different sized cores started being able to run simultaneously.

Apparently the smaller Tempest core from the A12 is being used as the main core in the new Apple Watch. Talk about, uh, synergy.

Yeah. The S4 Apple watch SoC is actually using Two Tempest (LITTLE) cores.

Pretty awesome, if you think about It. That little watch is more powerful than it appears.

I'm pretty skeptical of this. 3 W mobile CPU outperforms 165 W desktop CPU? If it was the case, they would have those things in macbooks already, and would boast tens of hours from one charge... If you go to the article that is mentioned in "recent figures", they tested single-core performance. So it was single core of a 28-core CPU (with boost though). It's still impressive, but not unbelievable.

It's a 165 watt desktop CPU running a single-threaded integer benchmark that doesn't use the other 27 cores, doesn't fully utilize the 6 memory channels or 40 PCIe lanes, etc. Also, you're comparing TDP to average power. Skylake won't come close to the TDP running an integer benchmark (it's the AVX units that really suck power).

Amber Lake Y (a 2-core part with 4.2 GHz turbo) should achieve similar performance on a single-threaded integer benchmark like that, and it's a 5W part. Still, that's incredibly impressive on Apple's part.

I'd consider it feasible-- in my testing, my iPhone XS outperforms my mid-2017 MacBook Pro (with i7-7820HQ [1]) in single-core benchmarking with Geekbench 4. Not by much, but it does manage a slight lead: https://browser.geekbench.com/v4/cpu/compare/10006209?baseli...

Granted, this probably wouldn't hold up for a more sustained workload since the fanless iPhone's going to hit its thermal limits much more quickly than the full-size laptop, but it's pretty impressive that they're edging this close to Intel anyways (in a much smaller, much lower-power package). It'll be really interesting to see what happens with the more unconstrained A12X that'll likely go into the iPad Pro this year.

[1] https://ark.intel.com/products/97496/Intel-Core-i7-7820HQ-Pr...

Did you run it on MacBook when it was on battery? Might not get to the max TDP in that case. So it would be like 30W vs 4W, but it's also 14nm vs 7nm.

Anyway though, it's still _very_ impressive from Apple, but much more believable. Looks like they've got very good engineers working on it. I hope they can scale it to the laptop requirements.

> Did you run it on MacBook when it was on battery?

Well, if the iPhone was also on a battery, that was a fair comparison.

I think that one was run on AC, but it was a while back so I can't be 100% sure. I keep meaning to re-run it to make sure, but haven't gotten to it yet.

I also have benchmarks of Haswell-era hardware which the XS is also very competitive against-- including my personal desktop, which the XS beats in single-core and matches for multi-core. But that one's probably not an entirely fair comparison, given that it's older hardware and the machine wasn't intended to be particularly high-end when I built it.

I'd like to see some more comprehensive benchmarking against Intel... I'd be inclined to do some myself, but I don't have a good benchmarking suite that isn't Geekbench. If only the SPEC benchmarks weren't so stinking expensive...

I'm also somewhat skeptical, but it's worth noting that "Skylake" covers a lot of ground[1], and the mobile Skylake processors use 15W (9.5W when in low power processing mode)[2]. Still a matter of multiples, but 5x instead of 55x.

1: https://ark.intel.com/products/codename/37572/Skylake

2: https://ark.intel.com/products/91169/Intel-Core-i7-6660U-Pro...

They talk about specific CPU there, Xeon 8176. That's why phrase from the article "we see that the A12 outperforms a Skylake CPU" is very sensationalist.

Ah, I see what's going on. The Skylake numbers are for a single threaded test. Assuming the A12 ones are as well, what's being said (poorly) is that the single threaded performance of Apple's A12 is better than the single threaded performance of Intel's Skylake Xeon 8176.

Of course there's a different number of cores in each, and the performance of multi-core code will likely differ based on technologies (memory architecture, hyperthreading vs real cores, etc).

If it's accurate, it is still fairly impressive (just not as crazy as it first sounds).

You're correct.

Anandtech is using SpecInt/FP numbers to compare Single core IPC between one A12 BIG core, and one Skylake server CPU core. Per-core power consumption on that Intel chip is 165/28 = 5.89W.

Apple is pushing similar IPC at 3W instead of 5.89W In that context, it's not so unbelievable, but it's still really fucking impressive.

Mobile skylake doesn't have 15w as the tdp limit, actual engineering docs sold under nda list real values.

This is the point where I mention that AMD and ARM documented TDP is max, where intels is an operational average.

Source: infrastructure operations for a large company using AMD, Intel and ARM CPUs.

SpecInt/FP is comparing Single core IPC between one A12 BIG core, and one Skylake server CPU core.

Per-core power consumption on that Intel chip is 165/28 = 5.89W.

Apple is pushing similar IPC at 3W instead of 5.89W. In that context, it's not so unbelievable. But it's still really fucking impressive.

Isn't it the display that eats most of the battery, not the CPU?

I think it is radio, especially if you are outside the city where the cell towers are far away from each other.

Depends. If the CPU runs at 100% utilization on all cores, I’ve seen battery life drop from 5 to 1 hour (for example).

The reason they don't have them in laptops is not because of performance, it's because they don't have an OS. A laptop that only runs iOS is called an iPad and it doesn't have the software I want to run.

I am quite sure that macOS runs quite fine on an A12. Like back in the PowerPC times, where Apple would build OS X always on x86 chips too, they certainly keep testing macOS on ARM hardware. Even if they don't have any imminent plans to release an ARM based Mac, it is a great way of testing your software for portability.

Of course, but without any software. Microsoft has done some interesting stuff in that regard so maybe there is an easier transition than getting all third party developers on board, but it's a tough nut to crack.

Apple have experience of that, and I mean a lot. The first C compiler I used was the cheap student version of Metrowerks that only compiled to 68K, I was on a PowerPC machine at the time. Mac OS X had a Mac OS classic compatibility mode until around the time the Intel processors replaced PowerPC ones, and early Intel machines could run apps compiled for PPC.

Apple is going to have a good ecosystem of ARM software very soon. The whole reason they pushed their Marzipan initiative with the ability to seamlessly port iOS (ARM) apps to macOS (x86) is so that they can also do (iOS >>> macOS) on ARM.

By asking developers to port from iOS apps, they have the side effect of getting developers to implement "fully featured" apps, just with a mobile UI. Adobe has already promised they'll be doing this with photoshop and Illustrator, releasing in 2019.

Others will follow suit. Not only will this strengthen the feature set of iOS apps, but it will strengthen the iPad Pro app ecosystem, and apple will simply link the iOS and Mac App Stores, allowing for "universal" apps.

Marzipan will translate them to x86/64 and later when Apple switches to ARM, Marzipan will simply keep these apps on ARM, and change their UI to match.

If you've been watching Apple closely over the last few years, you'll notice that they're consolidating their API's across macS and iOS. Things like Metal and AppKit and UIKit and so on are beginning to converge. This is all setting the stage for Marzipan.

They announced the "first half" of what they intend during the WWDC 2018 keynote. But the second half will be announced only after they reveal ARM Macs, and then "surprise - all the iOS apps you see now are available here, too. They automagically get UI adaptations for mac!"

Yes, the challenge is the software. Not so much the native Mac software created with XCode. That would probably rebuild at a button press. The elephant in the room are VMs. Many Mac users (for example me) are running x86 based operation systems inside VMs on their Mac. This wouldn't work without a dynamic translation like Rosetta. But as Apple is making their own processors, I would consider it a possiblity that they add some hardware accelerators for a Rosetta style software to their chips.

While I understand running VMs are important for you and many developers, people like us who want to do this have to be a miniscule share of Apple's market. I wouldn't bet on them spending much time on it.

You are right that developers are only a small part by volume. However they are the most important users in an environment, as they the ones creating new applications. Same with other power users, who require a VM. Driving them away from the platform could have a much larger net effect. That Macs got much more popular after the switch to x86 hints at that the compatibility isn't quite unimportant.

So I do think that Apple is probably going to ARM, but they need to be very careful in the steps they are taking and they need to have a really compelling offer so that users are willing to go through the transition.

My point was that if they had this abysmal advantage now, it means they would see it coming several years prior and would work on OS and anything else it needed.

But it's not the point, main point is that I think statement that current A12 is close (to whatever reasonable %) to top desktop chips is simply not true. I didn't see any concrete evidence of it, and everyone just seem to like this idea, that's why it was talked much lately. I'm all for great advantages and all that, but I still must use common sense, 3-4W vs 150+ W - that's too much.

People aren't comparing 3W mobile chips to 150W desktop CPUs. They're comparing like for like in terms of perf/watt. Skylake isn't a particular SKU, it's a micoarchitecture generation, from which Intel can scale the design from low power 10W parts to power hungry server chips.

Article reference specific CPU - Xeon 8176. It has 165W TDP.

SpecInt/FP is comparing Single core IPC between one A12 BIG core, and one Skylake server CPU core.

Per-core power consumption on that Intel chip is 165W/28cores = 5.89W

Apple is pushing similar IPC at 3W instead of 5.89W. In that context, it's not so unbelievable, and it's still really fucking impressive.

Yeah, I know, I wrote about it in my original comment, you didn't have to tell me that twice :)

And that's my point, Apple already has very impressive thing going on, no need to smudge it with sensationalists statements.

Apple has successfully changed architectures before, and has been developing an ARM version of Mac OS for a while. I am sure that it doesn't have the same level of refinement, it is might not be that far off.

1. It probably helps if you don’t have to accommodate 30 years of legacy.

2. The mobile phone industry is fiercely competitive while the PC industry has Intel and their near-monopoly.

Legacy isn’t a good explanation because the instruction decoder that turns x86 instructions into microcode is basically a fixed cost and makes up a tiny portion of the silicon. Clearly there is room for innovation in the x86 sphere as well because the AMD Zen architecture was created brand new from scratch and has become very competitive.

The more likely explanation is that there is an exponential difficulty curve in CPU optimization and performance that Apple and their admittedly talented engineers are climbing very quickly.

When an Ax chip hits true desktop performance sooner than later, the limit of physics is going to slow down Apple just like it has Intel and AMD. There isn’t any magic they are going to be able to do that suddenly gives them a 2x IPC gain or huge frequency advantage.

Well my Intel chip needs a fan and my iPhone doesn’t. So it’s not an apples-to-apples comparison (no pun intended, please nobody reply with “I see what you did there”).

The fact that someone can come up with a chip of similar performance to Intel’s isn’t that amazing. The fact that they can do so without needing a fan is.

I agree— though it’s perhaps worth mentioning that a fan on the phones was used for the SPEC2006 benchmarks:

> The performance measurement was run in a synthetic environment (read: bench fan cooling the phones) where we assured thermals wouldn’t be an issue for the 1-2 hours it takes to complete a full suite run.

Still darn impressive, as in the real world you won’t be doing anything with your phone that maxes out the SoC for 2 hours straight.

The performance comparisons are not remotely apples to apples to begin with. All that's been established is that an Ax chip is entering the realm of desktop performance. Under all the workloads you'd benchmark a regular desktop chip on (FPU, AVX, etc etc) the Ax chips are still going to be far behind at the TDP and frequency they run at.

If you could pump up the TDP and frequency of the Ax chips to actual desktop class and throw a heatsink on so you don't get thermally throttled, I am sure Apple's chips would be competitive.

It's an inevitability that we'll see a desktop-class Ax powered Macbook at some point soon-ish. When that happens we can start making legitimate comparisons.

Due to all that legacy the instruction decoder has a lot less opportunity for optimizations.

There may be a physical limit to the size of transistors but the limit of computing performance is not ‘what Intel is doing’. There’s more to performance than instructions per second or clock frequencies. In fact these don’t lead to the same performance at all as ARM is RISC and x64 is CISC.

That RISC and CISC, nobody mentioned until yours. Why are people only caring about Perf/watt or the SoC beating out the ST perf of the Skylake uarch. RISC based processor with that level of cache is always having an advantage due to the Load and Store at register level (esp in this A12 case of having HUGE L$) vs the CISC's more transistor level perf + DRAM to exec instructions, causing higher TDP requirements PLUS the multi instruction execution parallelism capabilities. And entirely different code of AArch64 and x86_64.

Add the GPU and Direct X, and Encode/Decode functionality I know this is not the subject, the ARM processors simply cannot keep up. The author did also mention the limitations of the comparison also as mentioned above the Skylake Xeon is capable of running Hex channel ECC Memory with huge capacity that alone will outpace everything on ARM it's not a fair share of comparison at all. GB is an OS level optimized bench. There's simply no comparison point that exists.

Like I'd say to the author do a 4D render bench like Cinebench then we talk on this. He should at-least have mentioned the uarch part. Also on a side note, Apple underselling the chips ? That's really first, since Apple always over promises. Their latest meltdown on the MacBooks with low quality/specced VRM throttling the CFL machines is a joke.

That x86 support isn't free. Intel's SIMD implementations and new instructions only ever seem to target the high end; I don't think you could use AVX512 at low power. They also guarantee various things about memory ordering that ARM doesn't, and of course the page table, 16-bit boot modes, and everything are super complicated.

I would expect there is some legacy aspect to it, although Intel building the Itanium (no legacy restrictions) just made a box that used more watts per instruction than their 32 bit line and wasn't that much better off. That said, things like the 'half carry' flag, segment pointers, and a large power hungry FPU don't help the Intel story.

I expect the primary difference one will find is that Intel has more "large" cores in their chips (full fledged cores) and the comparable instructions have some "gotchas" (like the flag mention) which make them harder to retire early in the pipeline. Toshiba made a funny comment at one of the Microprocessor forums that for their microprocessors, if you only compare the instructions that both they and Intel implement, the Toshiba processor is better in every way.

What is clearly true though is that compatibility aside, ARM chips are becoming serious contenders in the desktop/laptop space sooner than I expected them too.

The question should be for how long? Can it outperform a Skylake for 10 seconds? 30 seconds? 1 minute? 15 minutes?

A burst for a few seconds is ok on a phone, but not on a server or workstation.

But that isn't really the question rather than the answer -- in other profiles the chip could have significantly better heat dissipation solutions and would have dramatically more headroom.

That's not really the question, because turbo boost exists on servers and workstations too. Intel chips will clock to 4.5GHz for a short while before throttling down to their base frequency once the thermal ceiling is hit.

Huuuiuuge caches, and optimisations for using them as efficiently as possible: probably async prefetch, possibly a semblance of maskable cache, very likely reordering for cache missing ops.

Here, the prehistoric nature of x86 actually drags it down because of poor predictability

x86-64 is decently predictable, probably the same as armv8. i386 wouldn't be because it has to spill to memory so much you can barely manage to rename those registers.