Did This Apple Inc. Supplier Fail to Deliver?

It's widely believed that Taiwan Semiconductor Manufacturing Company (TSM 1.17%), or TSMC for short, is the sole manufacturer of the A11 Bionic chip that's at the heart of the Apple (AAPL +2.24%) iPhone 8, iPhone 8 Plus, and iPhone X. That chip is believed to be manufactured using TSMC's new 10nm technology, the very same technology used to build Apple's impressive A10X Fusion chip that powers the current iPad Pro tablets, as well as the new Apple TV 4K.

Image source: Apple.

Moreover, Apple boasted during its Sept. 12 launch event  that the new chip delivers significantly improved performance over the older A10 Fusion processor found in the iPhone 7 and iPhone 7 Plus, added functionality, and meaningful gains in energy efficiency. All that is good stuff, and it's what one should expect from a new-and-improved processor.

Interestingly enough, I ran into a post by independent technology analyst Daniel Matte in which he asserts that, based on what Apple delivered with the A10X chip, TSMC's 10nm technology is "really bad." While I'd disagree with Matte's characterization that it's "really bad," I think that all the chips we've seen using it highlight something worth exploring in a bit more detail. Allow me to explain.

It's dense, but does it move performance forward?

A new chip-manufacturing technology should aim to do the following:

1. Improve performance and power efficiency compared to prior-generation technologies.
2. Significantly reduce the area that an individual transistor (chips are made up of transistors) takes up, allowing for chip designers to either add more features while keeping chip area in check, or to maintain roughly similar features while reducing chip size -- and ideally in doing so, reduce cost.

TSMC's 10nm technology very clearly delivers on the second goal: The Apple A10X occupies a space of roughly 96 square millimeters, per TechInsights, while packing in more transistors than were included in the 147-square millimeter Apple A9X chip, which was manufactured in TSMC's 16nm technology. It's the peak transistor performance that appears not to have improved much, if at all.

To my knowledge, there are four chips in production using TSMC's 10nm technology: MediaTek's Helio X30, Huawei's Kirin 970, Apple's A10X Fusion, and Apple's A11 Bionic.

In the table below, I show the CPU frequencies of three of those chips and their direct predecessors (changes in CPU frequency generation-over-generation can certainly help us understand the magnitude of chip manufacturing technology performance improvements), which are all manufactured using TSMC's 16nm technology (I left out the MediaTek Helio X30 because it doesn't have a direct predecessor manufactured in 16nm technology):

Data sources: AnandTech, MacRumors, John Poole. 

At a glance, it looks like TSMC's 10nm technology doesn't deliver much of a frequency uplift compared to what its 16nm technology could deliver. The comparison of the Kirin 960 and the Kirin 970 is probably the most apples-to-apples comparison as the cores are identical, just manufactured in the different technologies. And the frequencies of the "big" and "little" cores in the 960 and 970, respectively, appear virtually identical.

As far as the Apple chips go, I'd say the A10 Fusion to A10X Fusion comparison is the most telling as the CPU cores are the same, though the A10X Fusion has more of them. Peak frequency of the high-performance cores in the A10X Fusion isn't up much from the peak frequency of the high-performance cores in the A10 Fusion.

Looking at the A10 Fusion to the A11 Bionic transition, Apple seems to have been able to push peak frequency up a little bit higher on a CPU core that, based on the Geekbench 4 CPU scores that have been published, performs more work per clock cycle than the high-performance CPU cores in the A10 Fusion could.

First impressions

It appears that TSMC's 10nm technology doesn't bring much, if any, gain in peak transistor performance/frequency capability over the older, more mature 16nm technology. Instead, the focus of the 10nm technology seems to have been on dramatically shrinking the size of the transistors to allow for chip designers to pack their chips with more stuff. This is something that's very valuable for mobile chips, as much of the "secret sauce" in a mobile processor lies outside of the CPU cores.

The challenges associated with such area reductions are far from trivial, and significant engineering needs to go into shrinking transistors. Additionally, since chip designers are naturally going to pack more transistors into their chips, those transistors need to consume less power so that overall chip power consumption doesn't get out of control.

It would seem, then, that TSMC delivered what it needed to: a 10nm technology that dramatically improves on transistor density and doesn't force a regression in peak frequency capability compared to 16nm. That's hardly a "bad" technology.