Apple’s “October Surprise”: the M3 SoC Family and the A17 Bionic Reprise

The meaning of the term “October Surprise” may not be immediately familiar to some of you, especially those outside the United States. Here’s how Wikipedia describes it:

In the politics of the United States, an October surprise is a news event that may influence the outcome of an upcoming November election (particularly one for the presidency), whether deliberately planned or spontaneously occurring. Because the date for national elections (as well as many state and local elections) is in early November, events that take place in October have greater potential to influence the decisions of prospective voters and allow less time to take remedial action; thus, relatively last-minute news stories could either change the course of an election or reinforce the inevitable. The term “October surprise” was coined by William Casey when he served as campaign manager of Ronald Reagan’s 1980 presidential campaign. However, there were October election-upending events that predated the coining of the term.

So why am I referring to Monday night’s Apple online (and prerecorded) “Scary Fast” launch event as an “October Surprise”?

Well, the first two reasons are literal:

The date of the event was October 30, 2023, and
Until a few days beforehand, nobody supposedly “in the know” believed (like last year) that Apple had anything else launch-wise planned for the remainder of 2023, save for maybe a handful of minor announcements delivered solely via press releases.

The third is more metaphorical. Obviously, nobody at Apple is a political candidate (…yet…). But the company more broadly has plenty of competitors, at least two of which—Intel in late September and Qualcomm just last week—made major processor announcements of their own. Was Apple’s event timing—or said another way, that of its competitors—coincidental, proactive, or reactive? We’ll likely never know, given all three companies’ tight-lipped cultures, but it’s still fun to speculate.

The SoCs

As the title of this post (not to mention the prior paragraph) note, Apple rolled out its M3 SoC Apple Silicon architecture for computers and high-end tablets this week. But, in a deviation from past-history trends, the company also unveiled M3’s first three product proliferations, along with initial systems based on them.

To understand why this was such a big deal, let’s focus first on the “A17 Bionic Reprise” portion of the title. When Apple rolled out the original M1 Apple Silicon SoC for computers in November 2020, as I mentioned at the time, it had notable architecture commonality with the A14 SoC for smartphones and tablets which had been unveiled two months earlier. This similarity isn’t a bad thing, mind you; I’d argue, in fact, that if the underlying architecture is capable of servicing the broader set of markets and applications, the resultant technology-sharing between the two SoC development groups can be an overall efficiency boost.

Granted, a SoC for computers and a SoC for smartphones differ in notable ways:

Varying numbers of both “efficiency” and “performance” processor cores, along with the on-chip cache that feeds them, and clock speeds
Varying GPU core counts and, again, clock speeds
Varying maximum capacities of system DRAM linked to both the CPU and GPU core constellations, along with other on-chip resources, as well as varying memory interface (and memory itself) bus widths and speeds, and
Varying system peripherals (including but not limited to the flash memory mass storage array) and interconnects between them and the SoC nexus

That all said, the various cores themselves are comparable between the two processor product lines, including special-purposes cores such as the neural engine and media engine.

Let’s go back to the M1. Recall that it ended up coming in four different proliferations:

The entry-level M1
The M1 Pro, with increased CPU and GPU core counts
The M1 Max, which kept the CPU core constellation the same but doubled up the graphics subsystem, and
The M1 Ultra, a two-die “chiplet” merging together two M1 Max chips with requisite doubling of various core counts, the maximum amount of system memory, and the like

But here’s the thing: it took a considerable amount of time—1.5 years—for Apple to roll out the entire M1 family from its A14 Bionic development starting point:

A14 Bionic (the M1 foundation): September 15, 2020
M1: November 10, 2020
M1 Pro and Max: October 18, 2021
M1 Ultra: March 8, 2022

Now let’s look at the M2 family, starting with its A15 Bionic SoC development foundation:

A15 Bionic: September 14, 2021
M2: June 24, 2022
M2 Pro and Max: January 17, 2023
M2 Ultra: June 13, 2023

Nearly two years’ total latency this time: nine months alone from the A15 to the M2.

I don’t yet know for sure, but for a variety of reasons (process lithography foundation, core mix and characteristics, etc.) I strongly suspect that the M3 chips are not based on the A16 SoC, which was released on September 7, 2022. Instead, I’m pretty confident in prognosticating that Apple went straight to the A17 Pro, unveiled just last month (as I write these words), on September 12 of this year, as their development foundation. Now look at the so-far rollout timeline for the M3 family—I think my reason for focusing on it will then be obvious:

A17 Pro: September 12, 2023
M3: October 30, 2023
M3 Pro and Max: October 30, 2023
M3 Ultra: TBD
M3 Extreme (a long-rumored four-Max-die high-end proliferation, which never ended up appearing in either the M1 or M2 generations): TBD (if at all)

Conceptually, the M3 flavors are reminiscent of their precursors, albeit with newer generations of various cores, along with a 3 nm fabrication process foundation. Apple seemingly focused the presentations’ “speaker notes” on M1 vs M3 two-generation-jump comparisons (not to mention x86 vs Apple Silicon comparisons), I suspect both because it made the comparative results more dramatic and because upgraders are more likely to come from a M1- vs M2-based system starting point. Here, for example, is an “efficiency” core generational comparison they shared:

Along with how the “performance” cores stack up:

Here’s an overall claimed CPU subsystem comparison:

Along with one for the GPU subsystem:

And here’s a broader comparison slide:

By claimed, I’m referring to the fact that Apple never (to the best of my recollection, at least) supplies supporting details that back up absolute-or-relative graphics and statements like these. What application(s) were being run? How does the system memory allocation compare? CPU and/or GPU clock speeds? Etc.? As usual, I’ll be dependent on third-party developers along with tech-savvy customers to fill in the gaps as well as point out issues that Apple’s glossed over.

Speaking of which…here’s one more comparative graph for you:

The M3’s 16-core neural engine (i.e., deep learning inference processing) subsystem is faster than it was in the previous generation. All well and good. But during the presentation, Apple claimed that it was capable of 18 TOPS peak performance. Up to now I’d been assuming, as you know from the reading you’ve already done here, that the M3 was a relatively straight-line derivation of the A17 Pro SoC architecture. But Apple claimed back in September that the A17 Pro’s neural engine ran at 35 TOPS. Waaa?

I see one (or multiple-in-combination) of (at least) three possibilities to explain this discrepancy:

The M3’s neural engine is an older or more generally simpler design than the one in the A17 Pro
The M3’s neural engine is under-clocked compared to the one in the A17 Pro
The M3’s neural engine’s performance was measured using a different data set (INT16 vs INT8, for example, or FLOAT vs INT) than what was used to benchmark the A17 Pro

Why would Apple do any of these things? Well, first off keep in mind that M-class SoCs are comparatively CPU- and GPU-core-rich compared to those in A-class SoCs, and those cores are perfectly capable of offloading (albeit perhaps less efficiently) some portion of the overall inference task. Note, too, that the inference applications that run on a computer are at least somewhat dissimilar to those on a smartphone. And finally, keep in mind that a computer either runs off AC or a beefier battery than what’s in a smartphone, as well as having beefier bidirectional broadband connectivity than a smartphone, so at least some amount of “offload to the cloud” is often feasible in the S-series scenarios.

There are a couple of other intriguing oddities that I want to point out. Take a look at this comparison of specifications for M2- and M3-series SoCs (I didn’t include the M2 Ultra, since it’s M3 sibling doesn’t yet exist “in the wild”, but simple extrapolation of the M2 Max dataset will get you in the M2 Ultra ballpark…to keep the table simple, I also didn’t show the M1 SoCs):

Spec

M2 Pro

M3 Pro

M2 Max

M3 Max

# of Performance CPU cores (max)

# of Efficiency CPU cores (max)

# of GPU cores (max)

Max # of displays

Neural engine

16-Core

15.8 TOPS

16-Core

18 TOPS

16-Core

15.8 TOPS

16-Core

18 TOPS

16-Core

15.8 TOPS

16-Core

18 TOPS

Memory and interface

LPDDR5-6400 DRAM,

8 x 16-bit channels (128-bit total),

100GB/sec peak bandwidth

LPDDR5-6400 DRAM,

8 x 16-bit channels (128-bit total),

100GB/sec peak bandwidth

LPDDR5-6400 DRAM,

16 x 16-bit channels (256-bit total),

200GB/sec peak bandwidth

LPDDR5-6400 DRAM,

12 x 16-bit channels (192-bit total),

150GB/sec peak bandwidth

LPDDR5-6400 DRAM,

32 x 16-bit channels (512-bit total),

400GB/sec peak bandwidth

LPDDR5-6400 DRAM,

32 x 16-bit channels (512-bit total),

400GB/sec peak bandwidth

Max memory capacity

24 GBytes

32 GBytes

36 GBytes

96 GBytes

128 GBytes

Transistor count

20 billion

25 billion

40 billion

37 billion

76 billion

96 billion

Lithography

5 nm

3 nm

5 nm

3 nm

5 nm

3 nm

Focus, in particular, on the darkened areas (duh, right?). Now, remember what I told you earlier about the M1 family, which also held true for the M2 family?

The entry-level Mx
The Mx Pro, with increased CPU and GPU core counts, and
The Mx Max, which kept the CPU core constellation the same but doubled up the graphics subsystem

Focusing first on the M2 Max versus M3 Max, the company’s done some reallocation of resources this time around. There are two fewer Performance cores this time, versus two more (smaller) Efficiency cores. The GPU core count has also dropped by 1, from 19 to 18. And maybe most notable of all, the system memory bus is 25% narrower on the M3 Max than it was on the M2 Max, translating to 25% lower peak bandwidth. The result? A net decrease of 3 billion transistors in the generational transition which, along with the more aggressive lithography, will translate into smaller (and at least theoretically cheaper, too, yield dependent) M3 Max die. Clearly, Apple’s relying on inherent performance improvements to the M3 architecture’s cores to counterbalance the lower Performance core count, and on on-SoC caching to counterbalance the decreased main memory bandwidth.

Re main memory, there’s one more interesting twist. Like its precursors, the M3 family implements a unified memory architecture, with CPU, GPU and other cores accessing the same DRAM array. And typically, the graphics (versus main memory) frame buffer allocation is fixed in size, even though in reality it varies over time as the user shifts between 2D and 3D content, for example, along with different resolutions, color bit-depths and other factors. With the M3 generation, Apple’s added an intelligent memory controller feature called “Dynamic Caching”, which accordingly varies the frame buffer size to optimize the amount of DRAM available for main memory applications, as well as the bandwidth to and from that DRAM:

Last, but not least, compare the M3 Pro to the M3 Max. Remember that in the past, the GPU core count has doubled, with CPU core counts staying the same. But this time, the Performance CPU core count has doubled (from 6 to 12), with the Efficiency CPU core count dropping by two (from 6 to 4). And the GPU core count has more than doubled (from 18 to 40). Apple clearly intended to craft a performance beast with the M3 Max and did extra design work to make it so.

The systems

Apple’s initial computers based on the M3 family are evolutionary in nature, and I’m therefore not going to spend much time discussing them compared the more revolutionary SoCs that power them. The M1 24” iMac was among the first introduced Apple Silicon-based systems, but Apple then skipped the M2 generation. The M3 iMac has more advanced Wi-Fi and Bluetooth facilities, commensurate with the newer chipset, but is otherwise identical to its predecessor.

Apple has also updated the 14” and 16” MacBook Pro laptops; the former starts with the M3 and can be upgraded to the M3 Pro, while the latter offers M3 Pro and M3 Max options. There’s a new (third) color scheme, space black (complete with matching cabling, but only available for M3 Pro and M3 Max system variants), to accompany traditional silver and space grey. Interestingly, as with the “base” versions of the M1 and M2 mac Mini, the “vanilla” M3 14” MacBook Pro only offers two Thunderbolt 4 ports and supports only one external display.

In related news, the 13” MacBook Pro (and the Touch Bar it integrates) have finally been “put out to pasture”. As I’ve mentioned before, the 13” MBP was becoming increasingly redundant with the 13” version of the MacBook Air; I’m surprised its demise took this long (even though I’m typing on one now).

And maybe the coolest thing of all about this week’s event (IMHO at least): the video footage was shot solely using the iPhone 15 Pro Max. Still accompanied by lots of fancy gimbals, supplemental lights, and other goodies, mind you, but still…wow. When I published my iPhone 15 family coverage back in September, I wasn’t yet aware that the Pro versions supported optional recording to USB-C-tethered external mass storage as well as LOG recording (not to mention wired Ethernet adapters and the like for all iPhone 15 models).

Having just crossed through 2,500 words, I’m going to wrap up for today. Let me know your thoughts on the new M3 SoCs and/or the initial systems containing them in the comments. As for me, I’m glad that my recent attempted eBay purchase of a M2 Pro Mac mini didn’t pan out; I’m now holding onto my money for the M3 Pro version that’s hopefully on the way eventually!

—Brian Dipert is the Editor-in-Chief of the Edge AI and Vision Alliance, and a Senior Analyst at BDTI and Editor-in-Chief of InsideDSP, the company’s online newsletter.

<!–
googletag.cmd.push(function() { googletag.display(‘div-gpt-ad-native’); });
–>

The post Apple’s “October Surprise”: the M3 SoC Family and the A17 Bionic Reprise appeared first on EDN.

Apple’s “October Surprise”: the M3 SoC Family and the A17 Bionic Reprise

Become a member

Become a member

Become a subscriber

Become a sponsor