Apple M1 to M6: how Apple Silicon evolved, and what each M, Pro, Max or Ultra is for

Six generations in six years, and four tiers in each. This is what actually separates an M from a Pro, a Max and an Ultra, and which one is worth paying for.

Part 2: the full evolution of the M series, from the 2020 M1 to the 2026 M6 and M5 Ultra generation

Continuing from part one: the M name alone tells you little

In part one we looked at why Apple moved away from Intel x86 processors to its own Arm-based Apple Silicon. Now it is time to follow how each generation developed. The most important rule to keep in mind throughout is this: M1, M2, M3, M4, M5 and M6 are not simply six consecutive clock-speed bumps of the same processor. From generation to generation, the CPU microarchitecture changed, and so did the GPU, the memory bandwidth, the media engine, the AI acceleration and the manufacturing process.

Understanding the product tiers is just as important. The base M is generally built for everyday and moderately demanding tasks. The Pro adds more CPU and GPU power, greater memory capacity and higher memory bandwidth. The Max typically offers an even stronger GPU and a larger memory system, which pushes it towards graphics, video, 3D and AI work. The Ultra, meanwhile, goes into the top-tier desktop Macs, where memory capacity and parallel computation matter most.

On 2 October 2026 the picture is a little unusual: Apple's current desktop line-up already includes the M6 and the M5 Ultra, while the current MacBook Pro family is built on M5, M5 Pro and M5 Max chips. In other words, the generation number and the product family do not always move forward in step. [15]

M1: the start of the revolution

The six Apple Silicon generations at a glance
The big steps are the manufacturing jumps and the AI hardware, not clock speed.

The M1 launched in November 2020 in the MacBook Air, the 13-inch MacBook Pro and the Mac mini. It was the first Mac chip that Apple designed specifically for this platform as its own Apple Silicon. The M1 was built on a 5-nanometre manufacturing process, contained around 16 billion transistors, and combined an 8-core CPU with up to an 8-core GPU and a 16-core Neural Engine. [2]

The main significance of the M1 was not that it instantly beat every PC in every category. The significance lay more in the overall picture. The MacBook Air gained a level of performance that we had previously only expected alongside much greater heat and power demands. The fanless Air showed that a notebook could house a powerful SoC while the system stayed essentially silent.

The M1's other big move was unified memory. Memory served as a shared resource for the CPU and GPU, and the entire chip could be treated as a single system. This is particularly beneficial for work that requires moving large data structures between the CPU and the GPU.

Alongside the base M1, it quickly became clear that Apple wanted to scale the family upward too. The M1 Pro and M1 Max arrived in 2021, followed by the M1 Ultra in 2022. This is how the product structure we still know today was born.

M1 Pro and M1 Max: the Pro and Max split is born

The M1 Pro was no longer simply a "faster M1". It offered up to 10 CPU cores, 16 GPU cores, up to 32 GB of unified memory and 200 GB/s of memory bandwidth. The M1 Max was built on the same 10-core CPU, but with up to 32 GPU cores, up to 64 GB of unified memory and 400 GB/s of bandwidth. [3]

This illustrates the Pro and Max philosophy well. On the CPU side the two chips sit closer together, while the Max opens up a bigger gap primarily on the GPU and memory subsystem side. This is why, for video, 3D or GPU-accelerated AI work, the Max's advantage can be far greater than in a task that relies almost entirely on the speed of one or two CPU cores.

The M1 Max also had a more capable media engine, with two ProRes accelerators. Apple specifically highlighted workflows such as multiple 4K ProRes streams, 8K ProRes and complex video post-production. Here, thanks to the fixed-function accelerators built into the chip, the goal was not to have the CPU calculate everything itself. [3]

The lesson here matters: when someone chooses an M1 Pro or M1 Max machine, they are not simply buying more "CPU power". A significant part of the premium is paid for the GPU, the memory capacity, the memory bandwidth and the media handling capability.

M1 Ultra: two Max chips in one system

The M1 Ultra arrived in 2022 for the Mac Studio. The Ultra's core idea was to take two M1 Max chips and connect them with a very high bandwidth link, using Apple's own UltraFusion packaging technology, so they could be used as a single system. The M1 Ultra offered up to a 20-core CPU, a 64-core GPU, 128 GB of unified memory and 800 GB/s of memory bandwidth. [4]

This was already a different league from a notebook chip. The Ultra was built for work where parallel computation and the amount of memory matter: professional video, large 3D scenes, scientific and development work, and large machine learning tasks.

UltraFusion matters because it showed that Apple did not want to rely solely on manufacturing a single, larger piece of silicon. A large chip can be more difficult from a physical and manufacturing standpoint, so the company was able to increase system-level performance through packaging technology as well.

M2: the first refinement, then the full build-out of the family

The M2 launched in 2022. It was built on a second-generation 5nm process and, according to Apple, brought an 18 percent faster CPU, a 35 percent stronger GPU, a 40 percent faster Neural Engine and 50 percent greater memory bandwidth compared with the M1. The available unified memory grew to 24 GB. [5]

The M2, then, did not represent the same dramatic leap as the M1. Instead, it refined and widened the direction set by the first generation: better performance on the CPU side, more power on the GPU side, more memory capacity and improved AI processing.

The real interest, again, lay in the Pro and Max variants. The M2 Pro had up to 12 CPU cores and 19 GPU cores, while the M2 Max received up to 12 CPU cores and 38 GPU cores. The M2 Max also offered up to 96 GB of unified memory and 400 GB/s of memory bandwidth. [6]

The M2 Ultra closed out the line-up in 2023. It joined two M2 Max chips together with UltraFusion, offering a 24-core CPU, up to a 76-core GPU, 192 GB of memory and 800 GB/s of bandwidth. This was a significant milestone, because 192 GB of unified memory enabled workflows where the bottleneck was not necessarily CPU speed, but the amount of data that the memory and the GPU could access together. [7]

M3: when the GPU architecture changed significantly too

The M3 launched in autumn 2023, and it marked an important technological shift for Apple Silicon. The M3 family was the first personal computer chip family that Apple introduced on a 3-nanometre process. The biggest new feature was not just the manufacturing process, but the new structure of the GPU.

Apple introduced Dynamic Caching, which allocates local memory in hardware, in real time, aiming to use only what is actually needed. Alongside this came hardware-accelerated ray tracing and mesh shading on the Mac. This mattered because, until then, ray tracing had been a strong differentiator particularly on PC-based NVIDIA and AMD systems. Apple now provided native hardware support for it within its own GPU architecture as well. [8]

The M3 Pro and M3 Max showed that Apple continued to scale the chip in multiple directions. The Max variant supported up to 128 GB of unified memory, and Apple specifically targeted it at machine learning developers, 3D artists and video professionals. [8]

No M3 Ultra was made. This is an important fact, because the numbering makes it easy to assume that every generation automatically has an Ultra. In the M series this is not a rule. Apple chooses different chip configurations for each generation and product.

M4: AI becomes a central focus

The M4 launched in 2024, first in the iPad Pro and then across the Mac line-up. It was built on a second-generation 3nm process. Apple added up to 10 CPU cores, a new GPU architecture, hardware ray tracing and a 16-core Neural Engine. The company specified the M4's Neural Engine at up to 38 TOPS, and separately highlighted the machine learning accelerators built into the CPU. [9]

With the M4 generation, AI stopped being a side feature and became one of the central goals of chip design. This matters because generative AI and local language models are not purely a GPU question. Performance is determined jointly by memory capacity and bandwidth, the GPU, the Neural Engine, the CPU and data movement.

The M4 Pro offered up to 14 CPU cores and 20 GPU cores, alongside 64 GB of unified memory and 273 GB/s of memory bandwidth. The M4 Max arrived with up to 16 CPU cores and 40 GPU cores, up to 128 GB of memory and 546 GB/s of bandwidth. [10]

This generation showed clearly that the Max is not necessarily better simply because it is twice as fast in every program. The larger GPU and the memory bandwidth become truly valuable only when the program can actually make use of them. In an office application, the difference between the M4 Max and the M4 is often barely noticeable. In a 3D render or with a large local AI model, however, it can be enormous.

No M4 Ultra appeared, though. This again confirms that Apple does not treat the Pro/Max/Ultra tiers mechanically in every generation.

M5: an increasing share of AI acceleration moves into the GPU

The M5 launched in October 2025, with the emphasis on AI even stronger. According to Apple, the M5, built on a third-generation 3nm process, brought a new 10-core GPU architecture with a Neural Accelerator in every GPU core. Apple claimed more than four times the peak GPU AI compute performance compared with the M4, while graphics performance also increased. The CPU has up to 10 cores, the unified memory bandwidth is 153 GB/s, and memory capacity is configurable up to 32 GB in the relevant models. [11]

Placing the Neural Accelerator inside the GPU cores is an important change of direction. It does not mean the Neural Engine loses its role. Rather, it means Apple offers several different computational paths for AI tasks. The GPU's massive parallel computation combined with the AI acceleration inside each core can be used more effectively for certain generative AI, image generation and LLM tasks.

The M5 Pro and M5 Max arrived in March 2026. For these, Apple used a design it calls Fusion Architecture, built on a two-die, high-speed interconnect. The M5 Pro and Max have up to an 18-core CPU, made up of 6 super cores and 12 performance cores. The Pro's GPU has up to 20 cores, the Max's up to 40 cores, and the Max's memory bandwidth grew to up to 614 GB/s, alongside 128 GB of unified memory. [12]

This is an interesting transition: instead of the classic "one giant monolithic die", Apple is scaling performance through more sophisticated packaging. This is also an important strategy from a manufacturing standpoint, and it shows clearly that in modern chip design, packaging can matter at least as much as the transistor itself.

M5 Ultra: the second level of scaling, now four dies

In August 2026, Apple introduced the M5 Ultra. This is a notable milestone, because according to Apple it is the first M-series system to use a quad-die, meaning four-die, design. The company achieved this with its next-generation UltraFusion technology.

The M5 Ultra offers up to 36 CPU cores and 80 GPU cores, alongside up to 512 GB of unified memory, with 1.2 TB/s of memory bandwidth. This is not simply "an M5 Max with bigger numbers". The Ultra's purpose is to serve professional tasks with extreme memory or parallel computation requirements. [13]

The enormous unified memory is particularly interesting from an AI standpoint. For a locally run model, what matters is not only how fast the GPU is, but also whether the weights the model needs actually fit into the available memory. If they do not fit, the system has to turn to another memory area or to storage, which is much slower.

This is why the Ultra's value, in certain cases, cannot be measured as "twice as fast as the Max". If the size limit on a workflow is memory, then a chip with more memory can handle a project that a smaller model simply cannot load in one go. This is the difference between a gain in speed and a gain in capability.

M6, 2026: a new CPU tier and two separate AI paths in a single base chip

The M6 arrived on 25 August 2026, first in the Mac mini, and became actually purchasable from 22 September 2026. Apple presented it, by its own account, as its first 2nm chip. The M6 uses a 12-core CPU: 2 super cores, 4 performance cores and 6 efficiency cores. The GPU has 12 cores, and every GPU core contains a Neural Accelerator. The M6 also received a Dual 16-core Neural Engine, so AI processing can take place through several separate hardware paths. Memory bandwidth, depending on configuration, is up to 170 GB/s. [13][14]

A 2nm manufacturing process alone does not guarantee that every application will be twice as fast. The technology node label is an increasingly imprecise shorthand for a simple "transistor density equals performance" formula. The real result comes from the microarchitecture, the cache, the memory, the GPU, the AI acceleration, the software and the power envelope, all acting together.

What is interesting, though, is that with the M6, Apple is turning even the base M tier into an increasingly AI-centred machine. The Mac mini M6 is no longer just a "general purpose computer". According to Apple's own measurements, it is also heavily optimised for running local LLMs, AI image processing and agentic AI tasks. Apple, for example, showed a significant LLM prompt processing speed-up under LM Studio compared with the M1. These are vendor tests, so they are not universal performance promises, but they show what Apple is optimising the platform for. [14]

It is also worth noting that no M6 Pro, M6 Max or M6 Ultra currently exists in the Mac line-up. In October 2026 the M6 appeared as the base chip in the Mac mini, while the more powerful machines are available in M5 Pro, M5 Max and M5 Ultra configurations. [15]

What do M, M Pro, M Max and M Ultra actually mean?

The four Apple Silicon tiers in the M5 generation
Between Pro and Max the CPU barely moves, while the GPU and the bandwidth double.

There is no single simple rule behind these names, but there is a practical model that works well.

The base M suits anyone who wants a fast, general-purpose machine. For office work, browsing, development, photo editing, lighter video work, running many applications at once and everyday AI tasks, this is usually the sensible starting point. Today's M6 is a good example of how the base M's performance now sits at a much higher level than the name alone would suggest.

The M Pro becomes interesting when longer, parallel, memory- or GPU-intensive work comes up regularly. Serious programming, compiling larger projects, running multiple virtual machines, large Photoshop files, more demanding video work, developer AI tools and professional multi-monitor use can all belong here. The Pro is generally not a "gamer's luxury", but the tier where the higher memory bandwidth and the extra CPU and GPU power can actually be put to use.

The M Max starts to make real sense once the GPU becomes the bottleneck. For 3D work, serious video post-production, large photography workloads, local AI and LLMs, large numbers of parallel GPU operations, and development tasks where GPU cores, memory capacity and memory bandwidth all matter at once, it is worth going for the Max.

The M Ultra, though, is already a different category. This is typically a desktop Mac Studio or similar top-end machine, aimed at AI models with enormous memory requirements, large 3D scenes, very many high-resolution video streams, scientific computation or professional content production. For an average user it is not pointless because it is a poor product, but because their work does not use its resources.

CPU, GPU or memory: which number matters more?

The most common buying mistake is to look at CPU cores first and then assume that a bigger number means a faster machine for every task. With Apple Silicon this is particularly misleading.

CPU cores matter most for general computation, code compilation, data processing, compression and a wide range of system tasks. A faster, newer CPU core can matter even on a single thread, because many applications do not scale perfectly across many cores.

GPU cores become more important when the software is capable of massive parallelisation. This includes 3D rendering, many video effects, certain AI models and numerous graphics operations. But software support is critical. A program that does not properly use Metal GPU acceleration or other acceleration will not perform miracles just because the machine has a Max chip in it.

Memory, however, is sometimes even more important. If the work fits in memory, everything is fast. If the system has to copy data too often or fall back on storage, the entire workflow can slow down. This is especially true for AI. A large model simply does not fit on a machine with little memory. In that case the question is not whether the CPU is 20 or 30 percent faster, but whether the model can be run locally at all.

This is why, when buying, instead of asking "M5 or M5 Max?" it is often a better question to ask "How much memory and what level of GPU performance do I actually need?"

AI and local LLMs: why is Apple's unified memory interesting?

The development of the M generations between 2024 and 2026 is hard to understand without mentioning AI. Apple placed an increasing number of AI acceleration features into the CPU, the GPU and the Neural Engine, while memory bandwidth also grew.

When running a local LLM, for example, the model's weights take up memory. A quantised model takes up less, a larger or less quantised model more. If the chip uses unified memory between the CPU and GPU, then the same pool of memory is accessible jointly to several processing units. This does not automatically make the Mac the best AI machine, but it is particularly interesting for tasks where large model size is the main constraint.

In Apple's M5 and M6 generations, the AI acceleration elements inside the GPU cores and the Neural Engine together form a multi-path acceleration model. Apple's own Foundation Models framework and its developer tools are aimed at letting developers use the CPU, GPU and Neural Engine for whichever task suits them. [11][13]

There is, however, an important counterpoint. The NVIDIA CUDA ecosystem still represents a huge advantage in the PC and server AI world. A great many research and development tools are built directly on CUDA. Local AI on Apple Silicon can therefore be a very good experience, but porting or running every AI project is not equally straightforward. Hardware alone does not solve ecosystem compatibility.

Which M chip should an average user choose?

For an average user, the choice between M, Pro, Max and Ultra should not be decided by prestige. Today's M-based Macs already offer a level of performance such that, for browsing, Office, communication, photo handling, general development and everyday creative work, the Pro or Max often does not visibly speed up the day-to-day experience.

For someone who is a developer and runs containers, virtual machines, databases and a large codebase at the same time, more memory and higher memory bandwidth matter sooner. For video work, the media engines and GPU performance matter more. In 3D work, the amount of GPU power and memory usually becomes the bottleneck. For local AI, memory capacity and GPU/AI acceleration matter together.

The current Mac mini M6 is a good example of how strong the base category can be. Apple pairs it with unified memory starting at 16 GB and configurable up to 32 GB, and bandwidth of up to 170 GB/s. For anyone doing more serious creative or technical work, the Mac mini M5 Pro is also available, with a configuration of up to 18 CPU cores and 20 GPU cores, 64 GB of unified memory and 307 GB/s of bandwidth. [15]

When is the Max worth it?

It is worth paying the premium for the Max when the bottleneck is genuinely on the GPU or memory side. An M5 Max's 18-core CPU configuration offers up to 40 GPU cores and 128 GB of unified memory, along with 614 GB/s of memory bandwidth. It is built for workflows where the graphics and AI load is heavy not occasionally, but regularly. [12]

A photographer who mainly edits JPEGs and a handful of RAW images will probably not make use of this headroom. A 3D artist rendering large scenes, an AI developer working with large local models, or a video professional handling a serious 4K/8K project, on the other hand, will more easily find tasks where the Max's resources genuinely save measurable time.

The best approach when buying, then, is not to look at the chip's name first, but at the longest and hardest step in your current workflow. If rendering, AI inference or video export constantly runs into a GPU limit, choose the Max. If the system's memory usage is the main problem, memory should be increased first. If the task is mostly single-threaded, even the largest Max chip does not guarantee a proportional speed-up.

And when does the Ultra make sense?

The Ultra is no longer a "better laptop" category. The Ultra's purpose is desktop-level work at very high capacity. With 512 GB of unified memory and 1.2 TB/s of memory bandwidth, the M5 Ultra targets problems with extreme memory and parallel computation requirements. [13]

In AI this may be the most interesting case. A large model may not only run faster, running it locally at all can become possible if the weights, the KV cache and the runtime environment fit into memory. The same applies to enormous 3D scenes and large video projects.

But the Ultra is not a better investment in every case. An office user who spends eight hours a day in a browser, Word, Excel and email uses only a fraction of the machine's resources. Choosing a Max or Ultra for this kind of work is technically possible, but little of the extra performance shows up in day-to-day use.

The biggest difference lies between "faster" and "capable of it". In certain cases an Ultra is valuable not because an export finishes 20 percent sooner, but because a given model, project or scene can be run locally on it at all.

Where the M6 currently sits in Apple's product line-up

At the start of October 2026, it is important to separate the chip generation from the device family. Based on Apple's current line-up, the M6 has not yet replaced the M5 generation everywhere. The 14-inch MacBook Pro is available with an M5, M5 Pro or M5 Max chip, and the 16-inch version with M5 Pro and M5 Max chips. [15]

The Mac mini, however, is already available with an M6 or M5 Pro system. The Mac Studio is built on M5 Max and M5 Ultra chips. This represents a transitional period: the newest-generation base M6 and the earlier, but higher-tier, M5 Pro/Max/Ultra sit side by side in the line-up.

It would therefore not be correct to say "the M6 is faster than the M5 Max in everything". These are different categories. An M6 brings a newer microarchitecture and modern AI/GPU technology, while an M5 Max delivers substantially more GPU cores and greater memory bandwidth. The specific workflow decides which resource is more useful.

According to Apple's announcement dated 22 September 2026, the new M6 Mac mini became actually available on that date, while the Mac Studio M5 Max and M5 Ultra models also went on sale then. [15]

Summary: how should we read the M numbers in 2026?

Several clear trends emerge from the development between M1 and M6. The first is integration: Apple keeps putting more functions into the SoC. The second is power efficiency: notebook performance is optimised together with power consumption. The third is AI: alongside the Neural Engine, more and more GPU-based AI acceleration has been added, particularly in the M5 and M6 generations. The fourth is memory: high bandwidth and unified memory are now one of the main determining factors for many professional tasks.

The other important trend is scaling. The base M is built for the wide majority, the Pro for more serious professional work, the Max for workflows with high GPU and memory demands, and the Ultra for top-tier tasks with extreme memory and parallel computation needs. The boundary, however, is not fixed, and can shift from generation to generation.

By 2026 it would be especially wrong to decide purely from the chip's name. A newer base M does not necessarily replace an older Max if the application uses the Max's larger GPU or memory bandwidth. At the same time, a newer-generation base chip can deliver performance, in a short, single-threaded task or through newer AI acceleration, that an older, larger chip may not necessarily match in the same way.

The evolution of Apple Silicon, then, is not really a simple staircase where every higher number is better at everything. It is more a case of development along several axes at once: CPU, GPU, memory, AI, media engine, manufacturing process and packaging. A good Mac choice requires weighing these against what the machine is actually going to be used for.

Once someone understands this logic, the names M1, M2, M3, M4, M5 and M6 stop looking like marketing labels and start looking like stages in a clearly traceable story of technical development.

Part 1 of this series explains why Apple left Intel and what unified memory actually does. If you are ready to choose a machine, our refurbished MacBooks are listed here, and we service Apple hardware in St Julian's.

Sources and verification points