Last week, Apple launched its M1 SoC. Built around the same A14 CPU as the latest iPhone, the M1 is the first non-x86 CPU architecture to challenge companies like AMD and Intel in well over a decade. Since then, there’s been a lot of back and forth about the Apple M1 chip, its relative comparison against AMD and Intel, and what various synthetic tests could (or couldn’t) tell us about relative performance. We’ve taken the position that the M1 itself isn’t going to collapse Intel and AMD’s market share, but that it represents a serious threat to both companies long-term.
So far, we’ve only discussed that theory with reference to recent synthetic benchmarks. Let’s now talk about the bigger picture.
How the M1 Threatens x86 Market Dominance
The reason the M1 is such a threat to Intel and AMD isn’t that Apple is suddenly going to consume the PC market. The Mac’s net market share gain is likely to be in the single digits, near-term.
The problem for AMD and Intel is what the M1 represents. For the first time in 15 years, a company without an x86 license is building a consumer microprocessor that plausibly competes with x86 chips. If Apple maintains or improves its position relative to Intel and AMD, other companies with ARM licenses are going to notice. Nvidia, which has purchased ARM pending regulator approval, is certainly going to notice. If ARM can outperform x86, the entire WinTel ecosystem is vulnerable in a way it hasn’t been since the dawn of personal computing.
Intel and AMD haven’t previously had to be concerned with what non-x86 CPU manufacturers were doing. If Apple’s M1 and inevitable follow-ups for different market segments start bleeding off x86 market share, everyone, including OEMs that currently build x86 PCs, is going to notice. Microsoft might have had Intel’s back once upon a time, but under Satya Nadella, the company has pivoted hard towards the cloud and cloud-based services. If Microsoft is willing to integrate Linux extensively into its own operating system, it’s not going to care if Windows runs on ARM or x86. The long-term threat to Intel and AMD is the loss of mindshare, margins, and market power if x86 is no longer seen as the automatic best processor choice. If that happens, it will be due to the actions of multiple firms, not just Apple. And no, despite Apple’s marketing, the current M1 isn’t faster than 98 percent of laptops. Fast, yes — but not that fast.
What makes the M1 more likely than any product we’ve seen before to kick start this kind of transformation? Here’s my reasoning.
Intel’s Pentium M Foreshadowed the Success of Core 2 Duo
In late 2004, a motherboard company named DFI brought a desktop board to market that was compatible with Intel’s Pentium M. This particular Pentium M was codenamed Dothan, and it was based on the Pentium III, not Intel’s Pentium 4 “Netburst” architecture. The 180nm P3 Coppermine begat the 130nm Tualatin, which begat Banias, which begat Dothan.
The DFI 855GME-MGF platform didn’t hold a candle to the FX and P4 platforms of the day, in terms of feature support. You can read my original review of the chip and platform here, via the Internet Archive, though you’ll need to use the “Next Page” buttons to navigate, and virtually all of the images are gone. One page that did survive, however, is the table of Unreal Tournament 2003 and 2004 results.
Let me re-familiarize you with what we’re seeing here: PM stands for Pentium M, tested at 2.13GHz and 1.7GHz. The FX-55 is AMD’s single-core Athlon 64 with dual-channel memory at 2.6GHz. The 3500+ is an Athlon 64 3500+ on dual-channel Socket 939 at 2.2GHz, while the P4 3.46EE, 3.8E, and 3.2E are all Prescott-based Pentium 4’s. These also have dual-channel RAM.
The DFI motherboard in question had virtually no overclocking support and there was no way to use anything other than the tiny, included fan and heatsink. In these results, the Pentium M is the only system limited to single-channel RAM, and the only system stuck using AGP. All of these problems notwithstanding, the Pentium M was a damn strong performer. It crosses swords with the P4 3.46EE and it beats the 3500+ (clocked 2.2GHz) despite the fact that the Athlon 64 3500+ has an integrated memory controller and dual-channel RAM.
To summarize its overall performance: Even when overclocked, Dothan didn’t win every benchmark and it dropped well behind the P4 and K8 in certain multimedia encoding workloads not shown above. Branchy tests that fit into its cache were a great test-case for Dothan; applications that relied on main memory bandwidth or that relied on SSE3 didn’t favor the chip. One of the places where the core stood out, however, was power. Here’s, erm, me, writing 16 years ago:
Drawing only 78W and idling at an amazing 25’C, Dothan is the coolest and most power-efficient CPU I’ve ever seen running as fast as it does; the 90nm 3500+ system draws significantly more. Even at full load, the system doesn’t break 100W… The 3500+ isn’t a power sucker; its 142W maximum is actually quite impressive compared to what the P4 draws these days… Prescott, in comparison, looks as though it hit every branch on its way out of the ugly tree.
Tech Report has an old review of this motherboard that survived with its images intact, so if you want to check their figures, you’ll find the data there. I’d like to quote from the conclusion of that review: “Overclocked to 2.4GHz on a 533MHz bus, though, the Pentium M gets downright scary, shadowing the performance of the Athlon 64 4000+ through many of our tests, including games.”
It wasn’t just shadowing. It was foreshadowing. Roughly 18 months later, Core 2 Duo, codenamed Conroe, arrived on the desktop and exiled AMD to 11 years in the wilderness. Dothan was the Silver Surfer to Conroe’s Galactus, if you want to use a Marvel metaphor.
There are multiple parallels between Dothan and the M1 today. Like the M1, Dothan didn’t win every benchmark. Like the M1, Dothan had certain limitations — it didn’t support SSE3, PCIe, or dual-channel RAM. It lacked the integrated memory controller AMD had, and it lacked the Hyper-Threading Prescott and Northwood had used to good effect.
Ultimately, none of that mattered compared with the Pentium M’s excellent performance per watt. In the long run, the design proved capable of outscaling any other competitor then on the market. Comparing the M1 against current CPUs from Intel and AMD, it’s striking how closely the patterns overlap.
Again, we see top-end x86 CPUs pulling ahead at higher TDPs in single-threaded benchmarks. The gap between x86 and M1 isn’t nearly big enough to offset the fact that top-end Intel and AMD chips carry 10-16 cores, while the M1 is a 4+4 chip. Given enough clock speed, thermal headroom, and cores, x86 carries the day. You could say exactly the same of the FX-55 and various Prescotts relative to the original Pentium M.
There’s a trend in the CPU industry that suggests high-efficiency CPUs with better performance per watt tend to scale to higher absolute performance levels than low-efficiency chips at weaker performance per watt. Because CPU power curves are not linear, the last few hundred MHz of clock is by far the most expensive in terms of power and thermals, and the curve gets steeper as you approach 5GHz. This means it can be more efficient to improve IPC by building additional on-die resources and reducing clock, even though the new functional units also cost power.
We don’t know enough about the M1 to make any predictions about how high it can clock, but the efficiency data we’ve already heard suggests Apple has at least some leeway to burn power to gain frequency. More likely, it’ll build a larger SoC, with more CPU cores. It’s not crazy to imagine a workstation-class Mx chip with larger caches and 16-32 big cores in the next 18-36 months. We could see 6-16 CPU core SoCs (derived either from FireStorm or the A15) within a year.
We can assume that TSMC’s 5nm node will continue to mature and that its clock speeds will improve. We can assume that Apple will consider all the standard levers and dials that CPU manufacturers typically tweak for improving performance in desktop power envelopes, like larger caches and higher clocks. The more efficient the M1 is at baseline, the more room Apple has to experiment with various methods of improving its performance. Right now, Apple has advantages in performance-per-watt and absolute power consumption in many applications, and that’s exactly where I’d want to be if I wanted to dethrone x86 over the long term.
Either way, the problem for both companies isn’t the M1 itself. It’s the M1’s big brothers that Apple will use for its higher-end products and the willingness of consumers and OEMs alike to move towards platforms with better battery life and higher performance.
AMD and Intel are reasonably well-positioned to fight the M1, though for different reasons. AMD is weaker in mobile than Intel is, but it’s got a killer narrative of consistent improvement carrying it from early 2017 through the present day. AMD’s Zen 3 mobile parts will improve its own performance against the M1 when they show up next year. Provided it continues executing into Zen 4, and Zen 4 keeps the improvements rolling, AMD is in little trouble.
Intel, meanwhile, is in pretty good shape at the moment. While it has faced its own manufacturing problems, Apple’s current M1 is focused on the mobile market, and Intel recently refreshed mobile with Tiger Lake. If Apple had shipped TGL hardware, Intel wouldn’t look as far back as it does relative to other Apple products. Short-term, the question is how the desktop/workstation version of the M1 might compare with high-end Xeons when Apple takes the chip into workstations. Longer-term, the big question for Intel and its remaining OEM partners is how the 7nm mobile transition goes.
If Intel can continue to deliver the kind of improvements it has picked up from the Ice Lake and now Tiger Lake transitions, it’ll have its own improvement story to talk about. The danger here is that 7nm is pushed back further or if the company is forced to outsource its leading-edge fab work. Intel isn’t going to try to win back Apple — that ship has sailed — but it will absolutely want to make certain OEMs have their own positive stories to tell around battery life and performance, even in comparison with ARM-powered systems.
The M1 isn’t magic. It isn’t faster than x86 just by virtue of being an ARM core; ARM’s standard licensed cores are nowhere near an x86 CPU’s performance. Apple has gotten itself into striking distance on x86 through years of careful work. Despite this, it’s not guaranteed that Intel and/or AMD will lose this fight.
But make no mistake — it’s going to be a fight. Nothing about the M1’s thermals, die size, power consumption, or core count suggests that Apple can’t keep scaling this chip. Nothing suggests it’s reached the end of its potential or that the core’s IPC can’t continue improving. Ordinarily, I’d say we should expect a yearly refresh, but since Apple is planning an entire transition to ARM next year, it might happen faster than that. By this time next year, however, the company should have a full lineup of products across desktops and laptops unless it delays its published transition timeline.
The last time we saw a comparison this lopsided between an up-and-coming big core competitor and two other high-performance, less-efficient micro-architectures, the newcomer and its children ruled the market for 11 years. Don’t assume the M1 is going to automatically win this fight. But keep in mind, this is Apple’s first salvo, not its last.
Intel’s Desktop TDPs No Longer Useful to Predict CPU Power Consumption
Intel's higher-end desktop CPU TDPs no longer communicate anything useful about the CPUs power consumption under load.
Review: The Oculus Quest 2 Could Be the Tipping Point for VR Mass Adoption
The Oculus Quest 2 is now available, and it's an improvement over the original in every way that matters. And yet, it's $100 less expensive than the last release. Having spent some time with the Quest 2, I believe we might look back on it as the headset that finally made VR accessible to mainstream consumers.
New Intel Rocket Lake Details: Backwards Compatible, Xe Graphics, Cypress Cove
Intel has released a bit more information about Rocket Lake and its 10nm CPU that's been back-ported to 14nm.
Intel’s Raja Koduri to Present at Samsung Foundry’s Upcoming Conference
Intel's Raja Koduri will speak at a Samsung foundry event this week — and that's not something that would happen if Intel didn't have something to say.