Apple Silicon: why Apple left Intel behind, and why ARM became so important

Apple did not simply swap Intel for another supplier. It moved the CPU, GPU, memory and AI hardware onto one chip, and that is what changed how a Mac behaves on battery.

Part 1. The fundamentals of Apple Silicon: ARM and x86, performance, power consumption, memory and the trade-offs

Introduction: the M1 was not simply a new processor

At the end of 2020, Apple introduced the first MacBook Air, MacBook Pro and Mac mini fitted with the M1 chip. At first glance, the change could look as if Apple had simply swapped Intel's processors for another manufacturer's solution. In reality, it was about far more than that. Apple carried over a system architecture into the Mac that it had spent more than a decade developing for the iPhone, the iPad and the Apple Watch. The goal was not just a faster CPU, but a reorganisation of how the whole computer works: CPU, GPU, memory, media engine, machine learning accelerators, security functions and I/O control, all inside a single, tightly integrated SoC.

Apple stated clearly back in June 2020 that one reason for the transition was to improve the performance-to-power ratio, and that the Mac and Apple's other products should share a common architecture. This also matters from a developer's point of view: the same foundations could be optimised for iPhone, iPad and Mac alike, while the Mac retained its own performance and expandability requirements. [1]

Why did Apple want its own chip?

The first important question is why an Intel processor was not enough. The answer has several parts. During the Intel-based Mac era, Apple relied on an external CPU supplier. That is not a problem in itself, but it limits how closely a hardware designer can align the CPU's capabilities with the rest of the system. Apple, by contrast, had long been designing its own processors for the iPhone and other devices. The M1 was essentially a far larger-scale version of that experience, tailored to the Mac.

With its own chip, Apple can decide for itself how to divide transistors between the CPU, the GPU, the Neural Engine, the media engines, memory control and other accelerators. This matters because modern computing tasks are increasingly not pure CPU tasks. In video encoding, for example, a dedicated hardware media engine can be orders of magnitude more efficient than when the same work is done by general-purpose CPU cores. The situation is similar for machine learning: certain operations are best handled by a separate accelerator.

According to Apple's 2020 announcement, one of the aims of Apple Silicon was precisely to improve the Mac's performance and its efficiency measured in watts, with greater GPU performance and technologies such as the Neural Engine. This does not mean that ARM is always inherently faster than x86. A chip's final performance is determined jointly by the microarchitecture, the clock speeds, the caches, the memory subsystem, the manufacturing process and the software. [1]

ARM and x86: what is the real difference?

ARM and x86 are, first and foremost, instruction set architectures, or ISAs. The ISA defines the rules by which software issues instructions to the processor. According to Intel's own description, x86 is an ISA that forms one of the fundamental architectures of the personal computer and server world. According to Arm's own documentation, the Arm CPU architecture is built on RISC principles. [18][19]

The classic explanation is that RISC is built on simpler, more regular instructions, while x86 historically inherited a more complex, variable-length instruction set. That is useful background, but today it should no longer be treated as if modern x86 processors simply work in a "CISC way" internally while ARM simply works in a "RISC way". High-performance CPUs of the current generation use very complex internal decoding, instruction scheduling, speculation and execution mechanisms. The external ISA and the internal microarchitecture are not the same thing.

The difference also shows up in compatibility. The machine code of a program compiled for x86-64 is not the same as that of a program compiled for ARM64. That is why Apple needed a transitional solution such as Rosetta 2 during the switch. Rosetta can translate the instructions of macOS applications built for x86_64 into a form that runs on Apple Silicon. Translation carries a cost, which is why a native ARM64 application remains the ideal. Apple's own developer documentation also states that native code generally delivers the best performance. [16]

This is one area where the advantage of Apple Silicon can only fully come through if the software is ready for it too. Between an old application optimised for Intel and a native Apple Silicon application, it is therefore not necessarily the processor alone that counts. The software translation layer itself is also part of the performance picture.

The key difference is not the letters "ARM", but system integration

Traditional PC memory layout versus Apple Silicon unified memory
Where the data sits decides how much copying the system has to do.

Many people conclude that Apple Silicon's success came simply from using ARM. That is too simplistic. The real advantage comes instead from the fact that Apple did not just bolt a separate CPU onto an old PC-style architecture, but designed the entire SoC as one.

In the M1, the CPU, the integrated GPU, the Neural Engine, the media engine and numerous system controllers all sit on the same die. This can reduce the cost of moving data around and allows the various subsystems to work together very closely. Apple calls this approach the Unified Memory Architecture.

In the traditional PC model, the CPU may use its own system memory while a discrete GPU uses separate VRAM. Because of this, the same data sometimes has to be copied between the CPU's memory and the GPU's memory in certain workflows. With Apple's unified memory approach, the CPU and the GPU can both work from the same large memory pool. This does not mean memory can be used without limit or at identical speed for every task, but the shared memory model can simplify a great deal of data movement and be more energy efficient.

This is particularly relevant for video and AI workflows. With a large graphics scene, a video project, or a local language model, it is not only the CPU's raw computing power that matters, but also how much data can be delivered quickly to the right processing units.

Why can an M1/M2/M3/M4/M5/M6 Mac stay fast for a long time on battery?

In notebooks, performance on its own is not enough. A processor can be very fast, but if that comes with high power consumption, strong heat generation and aggressive fan use, the user experience suffers. One of Apple Silicon's most important achievements is handling performance and power consumption together.

The M series uses a heterogeneous CPU design: high-performance cores and power-efficient cores work together. Apple's Developer documentation described the role of P-cores and E-cores as early as the M1 era. Simple background tasks do not always need to run on the highest-consumption cores, under heavier load, the performance cores can be brought in. [16]

The effect is striking in a notebook. The machine is not fast only when plugged into mains power. Under the right circumstances it also sustains high performance on battery. This is not equally true for every workload, and a long render or a large local AI model naturally represents a serious load, but the underlying principle matters: the chip, the operating system and power management are optimised together.

Power consumption also matters because less heat can mean less cooling is needed, and a given level of performance can be reached with less noise. One particularly telling example with the M1 MacBook Air was that it could make use of a significant share of the chip's performance even without a fan. When introducing the M1, Apple specifically highlighted the performance-per-watt ratio and long battery life. [2]

Where is Apple Silicon not necessarily better?

The three Apple Silicon trade-offs
Two of these three cannot be changed after you buy the machine.

Alongside its advantages, Apple Silicon also comes with real trade-offs. The first is expandability. Unified memory is part of the system directly connected to the SoC, so the memory capacity chosen at purchase typically cannot later be expanded the way it can in a classic desktop PC. As a result, for a given model, the amount of memory is a long-term decision.

The second is compatibility. Rosetta 2 made the transition much easier for a long time, but the translation layer is not magic. Specialised kernel extensions, virtualisation solutions, old plug-ins and certain low-level devices can be more problematic. Apple announced in 2026 that general-purpose support for Rosetta will end with macOS 27, while limited support for certain components tied to old, unmaintained games will remain. This means long-term compatibility with old Intel-only software is now a genuine consideration at purchase. [17]

The third is that integrated GPU performance is not equivalent, for every task, to that of a strong, discrete desktop graphics card. Apple has made great strides in graphics and hardware ray tracing capability, but support for games, 3D software and professional GPU applications still depends on the software. The largest M chips are impressive in many tasks, but a strong NVIDIA or AMD card in a Windows or Linux PC can still be in a different category for certain GPU-specialised work.

Finally, it is important to understand the question of price. Apple's integration is technically elegant, but memory and storage configurations can be expensive, and upgrading them afterwards is limited. For this reason, the same performance does not necessarily come at the same total cost as on a PC that can be partly upgraded by the user.

The key takeaway

The essence of the Apple Silicon story, then, is not that "ARM is better than x86". The real picture is more complex than that. Apple built a self-designed, ARM-based, heavily integrated SoC system in which the CPU, GPU, memory, AI acceleration, video processing and power management are all part of a single platform. This is a particularly strong combination for notebooks, because performance has to be optimised not in isolation, but together with power consumption and the thermal budget.

Intel's x86 world has not disappeared because of this, and x86 is not a "bad" architecture either. Intel's modern x86-64 processors use extremely advanced microarchitectures, with a broad ecosystem and enormous software compatibility. Today, the two worlds are better understood as a meeting of different design traditions and optimisation strategies.

The real question for a buyer, then, is not whether ARM or x86 appears in the specification, but how well a given computer fits the actual work at hand. And here, it is no longer enough to look only at the letter M. It matters a great deal whether a machine has an M, M Pro, M Max or M Ultra chip, how much unified memory comes with it, what GPU it has, what memory bandwidth, and what media accelerators.

Part 2 will therefore trace the development of Apple's chips from the M1 to the current M6, and separate out the base M, Pro, Max and Ultra tiers. The question will not be which is the "most powerful", but which technology is built for which task, and when the more expensive variant is worth it.

If you are weighing up a Mac right now, our refurbished MacBooks are listed here, and we also repair them in St Julian's. Part 2 of this series walks through every generation from the M1 to the M6.

Sources and verification points