“Apple Chipworks”, my nickname for one of the world’s most interesting chip designers, and holder of the “Overall Top-Performing Mobile SoC in Smartphones/Tablets” trophy since around 2012 or so, is a remarkable Mothership division.
They’ve clearly done it again with the Apple A11 Bionic, and according to the WSJ, they even have a custom-designed building within Apple Park — at their leader, Johny Srouji’s…let’s call it “purportedly insistent” behest.
A Neural Engine to support Face ID, a new GPU advertised as being a custom Apple design, breaking from the Imagination Tech roadmap…a person could write a ton about those 4.3 billion transistors, but it’s far above my pay grade, and not my focus for this post anyway. I’m taking a quick look at something a little more low-key…in particular, this:
Yep, I'm looking at the high-efficiency cores, which made their debut as a dual-core duo in the A10 Fusion, prompting many to think that Apple took a big.LITTLE page right out of ARM's playbook (even if that wasn't actually the case).
At a basic level (alas, that’s all I think I understand), the first-generation A10 efficiency cores operated as a "mini" dual-core CPU to low-power-compute through certain simpler tasks. Apple determined which dual-core CPU took on which tasks via its first-ever performance controller, which apparently didn’t need to exist until the low-power cores did.
But a High-Powered Turbocharger...
One of the notable features in A11 Bionic is the ability for iOS to use a single core, right on up to all six CPU cores simultaneously, thanks to the second-generation performance controller (Linked Source: Lance Ulanoff, Mashable).
And when called upon, the entirety of the A11 CPU array can, at least on paper, deliver mind-boggling (theoretical) multi-threaded performance gains over the A10. To say nothing of how well A11 Geekbenches on multicore against the closest SoC used by Android OEMs:
Why? Two reasons. First, the A11's ability to harness all six CPU cores at once, which the A10 Fusion was apparently not designed to do. Second, as nice as the A11's performance core gains are over A10, a 25% boost in performance core performance doesn't logically lead to a
Can You "Ballpark-Benchmark" an A11 Efficiency Core? Actually...Probably Yes
So I think we've established the four A11 efficiency cores act as potent turbochargers. And while there’s no concrete way to quantify their independent performance at this time (could the estimable John Poole and his team even "force" the A11 efficiency cores to benchmark on Geekbench 4 on their own?)…since Apple left a few clues over the past two iPhone keynotes, why not try?
I'll start with my key assumption: The A10 efficiency cores have a performance per watt of roughly 1/5 the A10 performance cores, because they actually use 1/5 the power.
So...how reasonable is this? I did check with chip-enthusiast Twitter pen pal, @TMFChipFool, and he told me that if anything, my wild guess was conservative. The reason: To him, Apple wouldn't go with low-power CPU cores that could completely take over certain computing functions of the iPhone unless they could provide better performance per watt than the high-power cores. When I think about it, this makes perfect sense; why build a set of power saver cores if your two main cores can do the exact same task at the same net power budget?
With this is mind, all we need to do is look at the "relative performance level" of the A10. And I've been tracking Apple's relative performance claims since it first started dabbling in custom silicon with the Apple A4 generation.
As you can see, based on Apple's performance claims...and year after year, Apple's performance boosts do seem to have been shown as legitimate...the A10 Fusion (high-power CPU) is around 23.8x faster than the A4.
So, take 1/5 of that, and you end up with a "performance score" of around 4.76. That's basically a "Apple A6 Plus" CPU score. And since A10, in my semi-power-user experience, at no point ever felt slower than the A9 in the iPhone 6s, even in Low Power Mode, I think it's quite reasonable to place Apple's A10 efficiency cores in the "A6+" compute power class.
Here's where it starts getting more interesting.
As you saw above, Apple's claimed a (up-to) 70% boost in A11 Bionic efficiency core performance. That raises the relative performance score from around 4.76 to (as high as) 8.1. Which is the dual-core Apple A7's relative performance level.
That's...only half of the truth, however.
See, Apple's performance claim for the efficiency cores is: "A11 Bionic has four efficiency cores that are up to 70 percent faster than the A10 Fusion chip".
For whatever reason, seemingly everyone in the tech and blogging sphere focused on the "up to 70% faster" and called it a day. And yet, there's an ambiguity hanging in the air.
FOUR efficiency cores.
That are up to 70 percent faster.
I decided to ask Apple myself whether
Interpretation 1: the claim meant that the four efficiency cores were collectively up to 70% more performant than the A10's two cores,
Interpretation 2: whether...as I was thinking, Apple meant that one efficiency core from the A11 Bionic was itself was up to 70% faster than one efficiency core from the A10 Fusion.
As it turns out, Apple's press contacts confirmed to me that Interpretation 2 was correct.
1) Apple dramatically speed-boosted the A11 Bionic's efficiency cores versus the A10 Fusion, and
2) There are twice as many "up to 70% faster" efficiency cores onboard compared to the A10.
Hence, what I call the "Double 'Seven". Because there are essentially two Apple A7 CPUs residing within the low-power compute profile of the A11 Bionic.
Since Apple's performance controller is obviously competent at getting the most out of the A11 Bionic from a multi-core standpoint, that also means that Apple's newest low-power cores are almost certainly faster at multi-core operations than the Apple A8. Which isn't much of a stretch, because when you refer back to the Geekbench Browser, the iPhone 6 Plus (Apple A8) has a multi-core score of 2423, only 13.2% better than the iPhone 5s (Apple A7) multi-core score of 2140.
If you take a little leap of faith and assume that the A11 quad-core efficiency CPU can multi-core process at least 50% better than the A7 on the iPhone 5s (you know, because twice as many cores), that places the "little" quad-core's (again, yes, theoretical) multi-core power at around the level of a Pixel C, a reasonably high-performance Google tablet from December 2015. Not too shabby.
More Than Megahertz
At the end of the day, though, this is more than about raw power. It's likely much more about power savings, defying the laws of battery physics and declining peak voltage for as long as technologically possible. Despite the additional system overhead that advancing mobile OSes almost certainly bring as the cost of progress, Apple Chipworks has more than managed to keep up by bringing the A11's efficiency cores to, theoretically, an A7 single-core level, and an A8+ multicore-level. This is turn opens up far more everyday tasks that the A11's low-power cores can smoothly handle in between incredibly demanding AR, VR, 3D gaming, image editing, and sooner than later...full-scale AI and ML applications. Someday, Apple's future efficiency cores may be relied on to perform a subset of those resource-intensive tasks.
And when you see one of the key claims from Apple's new custom mobile A11 GPU (A10 power at 50% energy consumption), Apple's strategy to maximize performance and battery life within the frustrating constraints of slowly-advancing battery technology becomes all the clearer.
Overall, seems Apple's silicon squadron is doing a very fine job pushing the performance envelope while expanding the envelope of low-power operation that users will never notice. Now Apple just has to work harder on the PR side of depleted Li-Ion battery reality.
 The Geekbench 4 math (I know, I'm going out a bit on a limb here) does check out. If you take the 5763 A10 Fusion multi-core score and boost it 25% (Apple's performance boost claim for A11), that amounts to around 7200. That would make the resulting "differential" in the A11 Bionic's Geekbench multi-score, which is nearly 3000 points higher at 10,111, "attributable" to the A11 quad-core efficiency cores. Consider further that Geekbench multi-core scores are not a linear multiple of [single-core score x number of cores] because physics and shared I/O architecture, and I think I just might have a reasonable handle on the A11 efficiency cores' multi-core processing prowess. (Don't even try to say that sentence ten times real fast.)