Another ARM Screed
Yes, it’s that time again. I will skip to the end right at the top and tell you what I think: ARM-based chips are going to eat away at the core businesses of Intel (INTC) and AMD (AMD). This is already happening, but it is in such slow motion, it’s hard to see unless you step back:
- Phase 1: Intel and AMD lose out on a huge opportunity and allow ARM SoCs to dominate smartphones and tablets. This is water under the bridge.
- Phase 2: ARM chips begin to show up in laptops, and even the data center. At first this is very small but the cost, customization and power-savings opportunities keep giving customers a reason to use ARM chips. This is where we are now.
- Phase 3: It’s less clear at AMD, but at Intel, laptops and desktops are still the biggest source of revenue. The data center is growing fast, and has higher margins. ARM chips are attacking from above and below in both their key markets. Eventually, the x86 platform becomes irrelevant.
This is a very long-term trend that began with the introduction of the first iPhone 13 years ago this month, and we are somewhere around halfway through it. But I believe we have passed the point of no return.
Since this is about a very long-term trend, there really is no investment thesis here, because Intel and AMD have years before this affects them in a big way, though Intel is about to lose about $5 billion in annual revenue from Apple (NASDAQ:AAPL). But it is clear to me that they are ignoring the risk, to their peril. They are so wrapped up in beating each other that they have failed to notice that someone else may beat them both. ARM is now one of the many tentacles of SoftBank (OTCPK:SFTBY) (OTCPK:SFTBF), and I cannot recommend investing in that.
For Apple, the Mac is now only a small part of their revenue line, which is dominated by iPhone and the fast-growing services and wearables categories. But any growth here would help offset their current issues in China, which I believe will continue.
So I will be back from time to time with updates on the progress of this slow-motion revolution when we hit a milestone. We have just passed one, and if rumors are correct, another is coming.
- Amazon (AMZN) turned on their ARM-based AWS instances and we are starting to get benchmarks. The results are not pretty for Intel and AMD.
- Apple looks to be announcing the upcoming 2021 ARM-based Macs at next week’s Worldwide Developers Conference (WWDC). This will be a huge milestone, but many questions and issues linger.
First a primer on ARM for those not familiar.
What is ARM?
ARM, originally Acorn RISC Machine, was a long-ago joint venture Apple put together, because they needed a low power CPU for the Newton, and no such thing existed then. Many years and a couple of leadership changes later, ARM CPUs wound up powering iPods, then iPhone, and then every other phone.
ARM does not make chips, but chipsets, instruction sets, and reference designs. It’s everything someone else needs to make a System on a Chip (SoC), which is a CPU, GPU and other parts that used to come in discreet packages, all on a single slice of silicon. A big brain in a tiny package.
For a company like Apple, who is obsessed with controlling their key technologies, this brings with it huge advantages. It allows them to custom design the SoCs for their phones to their needs. You can see their priorities in what they design:
- The CPU emphasizes performance at low power cost, rather than peak performance at maximum draw. Apple loves long battery life, which they consider the number one contributor to a good user experience in mobile.
- They take up a lot of space with the machine-learning cores, which they believe are key to their security and privacy goals, as well as their interest in AR.
- On the same note, the secure enclave lives on the A-series SoC, also a key to their security and privacy goals.
- The image signal processor, which is what makes what would be very ugly photos look fantastic, lives there too. Any look at iPhone marketing will tell you what they think of the camera as a source of customer satisfaction.
This is one company designing SoCs for a relatively small number of SKUs, so Apple can be very picky about what they invest in here.
Let’s contrast this with Qualcomm’s (NASDAQ:QCOM) flagship SoC. They have different priorities, because they are selling chips to a wide variety of phone makers, not phones to consumers.
- The CPU emphasizes the high performance cores that operate at peak draw. Battery life is not their problem; putting up big numbers in the marketing materials is more important.
- They have only recently put a lot of effort into the ML cores.
- There is lots of radio connectivity built into most of their SoCs, in contrast to Apple’s chips.
- They also have the image signal processor.
The point is not to say one is better or worse, just that each company got to build the SoC they wanted to build for their priorities and customers. Both SoCs use the same instruction set, but after that, they are pretty different. This is a huge advantage from ARM’s customers’ standpoint. But the most important advantage is power consumption.
Power consumption and its bastard child heat are the enemies of all hardware design. Everything you own could run faster, but would overheat quickly, so they get tuned down. When we talk about the data center, it becomes an even larger issue, as power is one of the primary variable costs. Big chips running hot and fast not only draw a lot of power, but require expensive environmental controls to keep the temperature and humidity down. It also will allow more density in existing data centers.
On mobile devices and now laptops, the advantages are obvious, with relatively high performance at very low power cost. Apple is right: battery life is the number one feature of an unwired device.
What ARM Brings to the Mac
In 2018, Apple announced their new iPad Pro sporting an A12X SoC. When it got out in the wild, everyone realized something amazing: the A12X was much faster than the Intel hardware in the MacBook Air announced at the same event. And of course it draws much less power.
The current A13 in the 2019 model-year iPhones has faster single core performance (what most of your apps use) than the Intel i9 in Apple’s current 16” MacBook Pro. And it costs so little that they can put it in a $399 iPhone SE, and still make an Apple gross margin on the phone.
If this is happening next week, as several Apple rumor sites have said, we can expect it to follow the patterns of Apple’s previous transitions: OS 9 to OS X, PowerPC to Intel, Intel to ARM for iPhone and iPad, and the elimination of 32-bit support in macOS.
- OS 9 to OS X was the roughest of them. The first two iterations of OS X were promising, but very buggy, and this left many sticking with OS 9, especially Adobe users, who did not see their apps updated for OS X for some time. The “Classic” emulation layer was slow, but many developers persisted with their OS 9 code for years after the transition. This one provides lessons in what not to do.
- After long ago switching to the PowerPC platform, Apple went back to Intel. This was a very smooth transition, which began with Apple having an Intel version of OS X running internally for 5 years before they announced it publicly. They announced it at WWDC 2005, and the first Intel Macs were available the following January. The transition was done by 2007 with the introduction of the first Mac Pro.
- The first part of their transition to ARM from the developers’ standpoint was the release of the first official iOS SDK with the iPhone 3G. This was fairly seamless, because Apple had based iOS around the same UNIX microkernel as OS X. There were such similarities, that even before the official release, hackers had created their own unofficial SDK, just modifying and adding to the OS X SDK slightly.
- Finally, we saw the elimination of support for 32-bit macOS apps with the most recent release, Catalina. Even though developers were warned years ago, still many got caught with their pants down. I bring this episode up for two reasons. First, eliminating 32-bit support is part of the ARM transition; there will never be a chip from Apple that supports it. Secondly, never discount the power of lazy.
As we look ahead, we can envision some amazing products we want to build for you, and we don’t know how to build them with the future PowerPC roadmap.
– Steve Jobs, 2005 WWDC
If he wants, Tim Cook can just crib from the master here. From the 2005 WWDC, where Jobs announced the Intel transition:
They can frame the decision in the exact way Jobs did 15 years ago:
- We have hit a wall with Intel.
- The key metric for us is performance per Watt, and ARM blows Intel away there.
The actual transition went incredibly smoothly, and began with every version of OS X being compiled for both platforms and running internally at Apple. Developers could compile their old code with a few modifications, and create a “universal binary,” with binaries for both platforms, which would install the proper version for the customers’ hardware. They also had a very fast emulation layer for older software. The whole thing was pretty invisible to customers.
This is clearly the model they would choose to follow. You can bank on the fact that ARM macOS has been running internally at Apple for some time now, as well as a version of their software development environment, Xcode. Moreover, they have likely been trimming out things from macOS (like 32-bit support), that will not work on ARM. I also imagine they will have Mac Minis fitted with ARM hardware available to the developers right away, so they can work in the correct environment, just as they had retrofitted iMacs in 2005.
So let’s add up what ARM does for Apple:
- It opens up huge design possibilities. I always thought the 2017 MacBook was a first crack at this – Jony Ive minimalism at its maximum. It was basically a pretty thin client with a relatively giant battery taking up almost all the internal space to provide all day battery life. I expect that the iMac will largely remain unchanged from the user’s perspective, but there will be all sorts of new laptop designs possible across their lines, with the power of a laptop and the battery life of an iPad.
- Apple loves control and this would give them total control over the most important thing: the future roadmap for all their CPUs.
- Like many Apple vendors, Intel has found Apple to be a demanding and difficult customer. Intel’s production issues pop up every few years, and it slows down the Mac cycle considerably every time. This is a source of great consternation internally at Apple.
- As I said, these SoC are much cheaper than the Intel hardware they are going to replace. The end result of this will be twofold, in my opinion. First, more gross margin. Apple loves gross margin. Who doesn’t? But also, we may see something like the iPhone SE in laptop form, like that 2017 MacBook, but cheaper and more powerful — an inexpensive (for Apple), very thin client with a big brain. This puts the Mac into new markets, where it will compete against PCs with far fewer capabilities.
The biggest downside from the user’s perspective is going to be for people who run a lot of virtual machines, like I do. Oh well, screwed again by progress.
The other big bit of ARM news is that Amazon has turned on their Graviton2 AWS instances, which they had announced several months ago. In the interim, we have seen some benchmarks from reliable sources, and it paints a very ugly picture for Intel and AMD. The headlines look like this:
Anandtech has a pretty wide ranging suite of benchmark tests for cloud compute, and the results match Amazon’s promise from the launch pretty closely – a 40% cost savings for the same job.
Customers are not going to flood these ARM instances all at once, and there is still a long road here. The three biggest hindrances are
- The general risk in switching to a new platform, whatever the savings.
- The dearth of ARM developers outside of mobile.
- The sunk costs of previous software development. We’ll get to the software aspect in a moment.
Also, a couple of very large caveats:
- Eventually, we are likely to see 6th generation Intel and AMD instances, and the new 6th generation ARM instances are being compared to the current 5th generation x86 instances.
- The ARM instances are not the most powerful thing AWS has on offer, and they are being compared to comparable Intel and AMD instances. More expensive and faster instances are available for the x86 platforms, and customers with demanding loads will continue to prefer these for now. But not everything is protein-folding simulations, and Graviton2 provides more than enough power for most of AWS’ customers.
Amazon’s moves here are important not just because of the innovation, but like Apple, they are a leader, and if they can be successful here, others will follow. Graviton2 is basically a slightly modified version of ARM’s reference design for their Neoverse data center cores, but other companies are pushing the envelope even farther.
Ampere, a private company, has been at this for a while, first with their eMag server chip. Like the original Graviton, it was a bit underwhelming. Their recently announced Altra is a different beast, literally pushing all of ARM’s Neoverse specs to their limits. This is a product aimed squarely at cloud providers and hyperscalers not named Amazon.
- 80 cores per socket running at 3 GHz, so a maximum of 160 cores in a server
- 4 TB of 72-bit DDR4 RAM per socket
- Up to 192 PCIe lanes in the 2 socket design
The numbers are just silly, but how it performs in the real world is still unknown. I’ll update when these things show up in the wild.
One final mention goes to Nuvia, another private company, but they have no announced products, and are being fairly secretive right now. It was founded by Gerard Williams, who formerly led Apple’s chip design unit. Naturally, Apple is suing Nuvia, and vice versa.
This all may be vaporware in the end, but Williams has assembled an all-star team of chip designers and ARM software developers, and there is a lot of buzz surrounding the company, and their very large ambitions for ARM in the data center. Again, nothing may come of it, but it looks very promising, just based on who is involved. Also, that Apple thought they were worth suing tells me they may have something good cooking.
Software Support Remains The Big Hurdle
The role that software can play in slowing this down or stopping it altogether came into sharp focus with Microsoft’s (NASDAQ:MSFT) Surface Pro X, an ARM laptop/tablet hybrid running Windows. It is undoubtedly the slickest hardware product Microsoft has ever produced. The screen is great, and battery life not exactly as promised but far better than my Dell laptop. It’s highly mobile, and it’s just a much more refined design than anything previously from the company. Windows 10 Home on ARM works just like the x86 version.
But just about all software runs in 32-bit emulation mode. Even Office does not yet have a native version; Microsoft could not be troubled to port their own flagship software to the new platform they are pushing. The experience was pretty disappointing.
But Apple, as usual, has much more control over this. In the first place, many of the apps used daily by Mac users are made by Apple, and ship with macOS. These are already running on ARM. Likely, the Pro Apps will have to wait.
Second, by 2023, every new Mac sold will sport ARM hardware, if the rumors are true. Microsoft certainly cannot dictate that to the PC makers. Developers will have to make the switch or lose out on new customers.
Finally, Apple owns the App Store, the number one place to sell Mac software. Apple will require all new apps to support both platforms come fall or winter, likely with the same universal binary structure we saw in the PowerPC to Intel transition.
That brings us to the data center, where the situation is much more complicated. Some companies, mostly larger ones, have years of sunk costs in x86 software, and the cost and trouble of porting may be prohibitive, no matter what the cost savings on the other end.
But the much larger group of AWS customers are smaller companies with smaller billing, and that adds up. Many of these companies are running web servers, game servers, database servers, app servers and other common network applications, for which there are plenty of open source packages available. This is the group of customers who will find ARM servers attractive at first.
So to sum up
- Third party software support is going to be difficult on Windows when even Microsoft can’t be troubled to port Office.
- But Apple can exert a lot of control here, and I expect the transition to be as invisible to the user as the PowerPC to Intel transition was.
- The situation is much more complex in the data center. Amazon’s best bet here in the short and medium term is the large number of small customers using open source packages to begin with.
Intel is much bigger than AMD, and they are also much more helpful in breaking out their numbers into segments, so we will stick with them. The first thing is that, although desktop/laptop (Intel’s “Client Group”) still remains the source of 57% of Intel’s top line, the data center has just passed the laptop.
Not only is the Data Center Group growing faster than the Client Group, but the margins are thicker, and that’s why it tends to get the most attention from analysts.
Turning it around to their customers, we see the PC manufacturers still dominate the top 10
The top 3 are as reported by Intel in their annual report. The rest are estimates by Bloomberg. Keep in mind that some of these, like WPG, are distributors, not end-customers. Chart © 2020 Trading Place Research
The top 4 are all PC makers that account for almost half of Intel’s revenue. This is largely laptop and desktop, but they also sell workstations using the data center chips.
The first important takeaway is that if Apple does indeed begin this transition next week, 7.4% of Intel’s revenue is going to get zeroed out, likely by the end of 2022. That is over $5 billion in annual revenue that will evaporate. Subtract that from whatever your best-case scenario is for Intel.
But what happens next? If Apple is able to make a very powerful 13” MacBook in the $600-700 range, a “MacBook SE,” the other PC makers will have to take notice. Those top 3 are 41% of Intel’s revenue, and they will have some big decisions to make, as Apple begins to roll out new ARM Mac models through 2021.
The big thing that will hold them back is the lack of a competitive ARM laptop platform from Qualcomm, the largest maker of ARM smartphone SoCs. Their 8cx compute platform is pretty underwhelming, even compared to Apple’s phone SoCs. That they don’t list the clock speed of the CPU in the marketing material should tell you everything you need to know.
About 85% of Intel’s Client Group revenue comes from their top 20 customers, but only about half their data center revenue comes from there, so it is much more diffuse. The “Next 10” wedge in the last pie chart is 1 distributor, Acer, and 8 data center customers such as Amazon, AT&T (T) and Facebook (FB).
So what happens here is going to depend largely on Amazon’s success. Alphabet (GOOG) (NASDAQ:GOOGL) and Microsoft will be watching closely, and if Amazon gains some traction, they will have to jump in. Unlike with the laptop, there may already be a robust platform for them to adopt, Ampere’s Altra.
So even more so than with laptops and desktops, this will be a game of follow the leader. The key question is whether the reduced costs of ARM servers can make up for all the other costs and risks involved with switching platforms. Enterprise customers tend to be risk-averse with these things, but they also like saving money. Stay tuned.
Amazon is one of the strangest companies of all time. Everyone thinks of them primarily as a retail company, and they are if you look at their top line.
AWS is still a small, but growing part of the extended Amazon family in this view. But operating profits tell a very different story.
I know of no other company where 12% of the revenue accounts for 63% of operating profits. It’s getting better, though. In 2017, AWS was only 10% of revenue, but 105% of operating profits, because the rest of Amazon operated at a loss that year.
The point here is not to trash the rest of Amazon, which I’m happy to do at another time, but to show you how important it is to the entire operation, which may have collapsed under its own weight at this point without AWS.
The only reason they can print that 5% consolidated profit margin is because of AWS.
So point number one is that Amazon is not going to do anything risky here and kill the golden goose. They invented cloud computing as we now know it, and remain the leader. This is already a big business, and it is growing rapidly. Though they have in the past with other projects, they did not enter into this lightly.
Whether they succeed is another story, but if they do, it will fund the profligacy of the rest of Amazon for years to come, and that is good news for Amazon investors. So what advantages does it provide?
- Like with Apple, it allows them to control their product roadmap. Down the road, I can see them designing chips that are optimized for unique workloads, which would provide a huge advantage over the competition. Intel and AMD could never do this for them.
- Smaller chips producing less heat means more per rack, and greater density throughout existing data centers. To add to Jobs’ metric, performance per Watt per cubic meter is what’s important.
- They are already the price leader in most ways and this will guarantee that continues.
Amazon’s big customers are unlikely to go for ARM servers right away, or even in the medium term. They have tons of sunk costs in the x86 platform, and are risk-averse with these sorts of things in any event.
The best chance they have for gaining traction in the near term is with small customers using open source packages, for which there is plenty to choose from. Also, if you were a startup developing a mobile app, and had a high cloud bill, developing on ARM end-to-end may seem like a natural and cost-effective solution.
For Apple, the Mac would have to see substantial growth to make a dent in their income statement.
The Mac is now only 10% of Apple’s revenue. To be clear, this is still $26 billion annually, and Mac’s best year ever in 2019. But even if these new lighter, cheaper, more powerful Macs with longer battery life became such a huge hit and that line went up by half, it would still be only a 5% bump on their consolidated top line.
But it can help with this:
Apple Annual Reports. Chart © 2020 Trading Place Research
After years of only going up, Apple’s revenue from China fell off in fiscal 2019, and did not recover with the 2019 iPhones released in September. In all, it amounts to about $8 billion in lost revenue. With current tensions, I don’t think they will be seeing that line pop back up any time soon.
Until the trade war, Chinese consumers were a huge part of Apple’s long-term plan, which has now been derailed by Trump. So any increases in the Mac line are sorely needed to offset declining Chinese revenue, though mostly this is happening through the growth of services and wearables.
The new long-term plan involves India as an escape hatch, but that’s an entirely different article.
The Slow Revolution
When Steve Jobs announced the first iPhone in June of 2007, this revolution began. At the time, Intel was the dominant platform and Microsoft had the dominant operating system. Thirteen years later, neither is true. There are many billions of Intel chips out there, but now hundreds of billions of ARM chips, and Google’s Android is by far the number one operating system. Intel’s is one of the great lost opportunities in the history of capitalism, matched by Microsoft’s.
But ARM is coming far more than that, and now we have two market leaders, Apple and Amazon, pushing hard in that direction. The next 13 years will be equally as interesting.
Disclosure: I am/we are long AAPL. I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.
Originally published on Seeking Alpha