In 2020, Apple made a switch in the company supplying the processors for its Mac computers, from Intel to Apple itself. The rise of homegrown silicon for Apple had begun more than a decade before with the iPhone 4. With the M1 chip, the silicon in its Macs was built in-house, too.
The new processors brought a huge improvement in performance, offering a big uptick in speed and, crucially, battery life.
Since then, there have been multiple upgrades, with the latest M5 chip appearing in the MacBook Pro and the iPad Pro. The iPad, uniquely, uses a mix of M and A chips in its range. I sat down to talk exclusively with Tom Boger, Apple’s Vice President of Mac Product Marketing and Tim Millet (pronounced Mill-ay), who is Vice President Platform Architecture.
First, I ask if that big uptick in performance with M1 was a one-off and we should expect smaller increases now.
“When the M1 first arrived, moving from the Intel architecture to Apple Silicon was a huge step change, and we were excited to put a Mac based on Apple silicon in the hands of our customers, and it was just radically different,” Boger says. But since then, we haven’t taken our foot off the gas. We continue to drive the architecture as hard as always. And M5 introduced another step change, and that’s from an AI standpoint. We brought neural accelerators to the GPU, so the AI compute with M5 when you compare it to M1, is six times greater, and four times the compute from M4. We’ve continued to push further on all the vectors in the chip. So the performance difference between now and then is pretty huge.”
“We built excellent systems for probably more than a decade with Intel, and moving to M1 was not just moving to a different chip, we were moving to a different architecture,” Millet says. ”Intel didn’t just build a CPU, they built a platform where they brought their pieces, and they worked with an ecosystem of partners to build their pieces. It was a super-successful platform for many years. But in the end, when we had built up enough capability with our chips for iPhone and then for iPad, we realized there was a different architecture there that was actually better suited to where we wanted our Mac to go.
“And this architecture was one with a focus on energy efficiency. It also had a focus on the unified memory, which is sharing a common pool of memory between all the major compute blocks We had built up this toolbox of capabilities through our iOS products, bringing them to the Mac, and creating a sense of wow. But it wasn’t day one and done. It was a desire to move the Mac up to the same state-of-the-art curve that we have for all our products. I think AI is going to be an area where the unified memory that we’ve introduced into these platforms is a perfect fit. Now, did we know that that was going to happen back in 2020, when we launched it? I think we were trusting our judgment that a unified memory platform is much more flexible.”
So, did Apple predict that memory-hungry AI demands or the needs of large-language models like ChatGPT were coming?
“When you combine pools of memory into a high capacity memory system, you create something somewhat interesting,” Millet says with a note of understatement. “And again, we were trusting our architectural judgement, that this is a better platform for where we want to take the Mac.”
Since the launch of the M1, M chips have appeared on the most powerful iPads, too. I’m interested in how this commonality came about. “It was a very natural thing. The transition of the Mac from the Intel architecture to the Apple Silicon platform brought the iPad and the Mac into a common technology pool. And if you look at kind of what the needs are that are common, it’s a huge overlap. You could ask how, as a chip designer, you could keep both customers happy. It’s because we co-design with our system partners, our software partners, our design partners, to make sure that every detail is taken care of.
“We scaled the neural accelerators from the A19 Pro chip in the iPhone 17 Pro and introduced them in M5, in a much bigger GPU appropriate for the MacBook Pro and the iPad Pro. We didn’t think, hey, this M1 is great. We should try it on an iPad. We plan years ahead. We know exactly where our chips are going, which is a huge benefit for my team to know exactly what our targets are. We can anticipate the product, and that’s where architecture really matters. You want to make sure you’re building a scalable, robust platform. We made investments in M1, even, to include the neural engine that we had introduced into the iPhone years before, and that has paid dividends, because all of our Apple Silicon Macs now have the ability to run on-device LLMs, at a scale, and at a performance level. And part of it is the neural engine.”
The neural engine was first introduced as “a way to extend camera processing, really, with a focus on computational photography,” Millet explains. “What we did that was interesting was we took the neural engine from being a component of the camera processor, extracted it and centralized it, so it became a first-class compute engine. But we felt like there was more opportunity to go beyond just photography, which is why we wanted to make sure it was available to any application. It was a little risky. I always have to be careful about introducing new hardware, because if we put it in and software doesn’t use it, we call that dark silicon. It’s transistors on the chip we paid for, but we’re not making great use of. I get in trouble for that.
“This was in 2017 that we released this in the iPhone. That same year, my team that built a neural engine.They pored over every interesting Machine Learning paper and saw this paper called “Attention Is All You Need”. This is the paper that led to the development of the Transformer model, which allowed all of these large-language models in the generative AI explosion. The neural engine came way before all that, in 2020, when we released M1, two years before the large language models showed up. M1 had support, really, first class support for running this new class of models, even before it was obvious why you should.”
In hindsight, Apple M1 chips, and the later M chips have been a big success. I wonder whether that looked like being the case before the M1 launched.
“Our mission is, build the best possible product that we can with the tools that are in front of us,” Millet says. “The Mac was in a great place when we introduced M1. We did not want to go out the door and miss something obvious. So, we spent a good amount of time thinking through all the different things that mattered to the Mac, really getting engaged with the system teams and the software teams. But when you got all of that input, it really did help solidify our confidence that this was going to, this was going to get out without a bump. We’re never anticipating huge success. We’re always worried about the thing we missed. The team is never resting on their laurels. They’re always focused on what comes next. Some of it is a judgement call, but it’s always about looking into the future. My team, we live three to four years into the future.”
Annoyingly, neither Millet or Boger will be drawn on what that future will look like. So, finally, I play devil’s advocate. There are plenty of people who bought Macs with M1 processor and don’t need to update their machines because they’re still so fast and effective. Did they do their jobs too well?
Boger is keen to answer. “First of all, we love that sentiment. We love the fact that people are so satisfied and happy. If they’re still happy on an M1 system, that’s great, because that’s our goal, right? We’re constantly trying to, number one, build the best product we possibly can, but number two, give our users the best experience possible. And if they, right now, are walking around with an M1 based Mac and they love it, and it’s meeting their needs, we’ve done our job.
“But then, you know, we haven’t rested on our laurels. If you look back, compare an M5 to M1, and you look at battery life, or CPU performance, GPU performance, AI performance. Now, the gains are starting to become pretty massive in the latest systems. It’s like we’re challenging ourselves and setting a high bar with each generation to see if we can create something that’s great at the time, but also relentlessly asking ourselves, how do we push things even further.”
Millet adds that he has “friends who are constantly upgrading because they need the compute. Folks who are working on media processing are now looking at M5 and saying we have enabled them to more easily extend applications with AI, which is magically faster than it was before. Or take metal-effects scaling, a tool that allows the GPU applications rendering for gaming to essentially render something at low resolution but use AI-driven scaling to make it look like high fidelity, much higher resolution. This used the neural engine.
“All of our engines are designed to operate in tandem. So I think we’re excited about making sure the Mac is never in the situation it was in 2020, where the state of the art was high, and the Mac was living on an architecture that was a step function below. And the idea that we would just let the technology curve move forward and leave the Mac behind doesn’t make any sense.”
