MATLAB System Requirements (2026)

Estimated reading time: 13 minutes

There’s a conversation I’ve had maybe fifty times now. Someone messages me a laptop listing, asks “will this run MATLAB?”, and the answer is always the same two words: yes, but.

Yes, it’ll install. Yes, it’ll open. Yes, you can plot a sine wave. But will it still feel usable in your third semester when the assignments get real, or six months into a project when your datasets stop being toy-sized? That’s a different question, and the official requirements page won’t answer it for you.

That page isn’t wrong. It’s just doing a different job. MathWorks publishes the floor — the point below which the software won’t function. Most people read it as a recommendation. It isn’t.

So let’s do both. The real numbers for MATLAB R2026a, laid out clean, followed by everything I’ve learned about what actually makes MATLAB feel fast versus what makes it feel like wading through wet cement.

Start here: the short version

If you skim nothing else, note the letters AVX2. There’s a countdown attached to those four characters and I’ll explain it shortly.

What MathWorks officially requires for R2026a

Straight numbers, no commentary. Commentary comes after.

Windows

macOS

Linux

Running something not on that list — Fedora, Arch, Mint, Pop!_OS? MathWorks says distributions with kernel 4.18 or later and glibc 2.28 or later will likely work, you just won’t get full support if something breaks. In practice that covers nearly every modern distro.

One more Linux note that saves people hours: MATLAB is validated against standard installations. Minimal or core images frequently lack packages the installer expects. If you’re working from a server image or a slim container base, that’s where the trouble starts.

Toolboxes that skip your platform

Check this before the OS decision hardens.

Missing on macOS: Data Acquisition Toolbox, GPU Coder, Deep Learning HDL Toolbox, HDL Verifier, Vision HDL Toolbox, Vehicle Network Toolbox, Simulink Real-Time, Simulink Code Inspector, SoC Blockset, Model-Based Calibration Toolbox, C2000 Microcontroller Blockset, RF PCB Toolbox, Signal Integrity Toolbox, Spreadsheet Link, Polyspace Client and Server for Ada, Polyspace Access, RoadRunner and its asset libraries.

Missing on Linux: Data Acquisition Toolbox, Spreadsheet Link.

Anyone doing hardware-in-the-loop work, automotive code generation, or FPGA verification just had their decision made for them. Windows, or a Windows virtual machine.

The AVX2 deadline nobody talks about

Tucked into a footnote on both the Windows and Linux pages is a line that matters more than anything else on those pages: a future release of MATLAB will require a processor with AVX2 instruction set support.

That’s a warning shot. AVX2 showed up in Intel’s Haswell generation around 2013 and AMD’s Excavator around 2015. Anything newer has it. Anything older doesn’t, and when MathWorks pulls that trigger, those machines are done — not slow, not degraded, just permanently stuck on their last compatible release.

Checking takes half a minute:

• Windows — open PowerShell, run Get-CimInstance Win32_Processor | Select-Object Name, then look the model up. Or install CPU-Z and read the Instructions line.
• Linux — grep -o avx2 /proc/cpuinfo | head -1. Output means you’re covered.
• Mac — doesn’t apply. Apple Silicon uses NEON.

This matters most for institutional buying. University labs and small firms keep machines for six, seven, eight years. A CPU without AVX2 purchased today has a shelf life measured in releases, not years.

Memory is where most people miscalculate

MATLAB stores numeric arrays in contiguous memory blocks. That single design detail explains most of the frustration people have with RAM.

A 10,000 × 10,000 double array is exactly 800 MB, and it needs 800 MB in one uninterrupted stretch. Now run an operation that creates a temporary copy. Then another. You’ve consumed 2.4 GB writing three lines of code that looked completely innocent.

This is why “out of memory” errors show up on datasets that sound tiny. It’s never just the data — it’s the data, its temporaries, the IDE, the browser you left open, and whatever the operating system decided to do in the background.

Numbers I’d actually stand behind:

That last line trips people constantly. Opening a pool of eight workers on a 16 GB laptop doesn’t give you eight times the speed — it gives you a machine that spends its life paging to disk.

Diagnostic worth memorizing: if MATLAB is crawling and CPU usage is low, you’re thrashing, not compute-bound. Low CPU plus terrible performance equals memory pressure, every single time. Check Task Manager on Windows, Activity Monitor on a Mac.

Processors: what helps, what’s marketing

MATLAB multithreads a lot of operations automatically — linear algebra, elementwise math, FFTs. But not everything is multithreaded, and the benefit varies enormously by algorithm.

Core count helps until it doesn’t. Four to eight cores is a genuine improvement you’ll feel. Sixteen to thirty-two mostly pays off if you’re deliberately using Parallel Computing Toolbox and structuring work as parfor or parfeval.

Hyper-threading barely moves the needle. Virtual cores appear in Task Manager and cause endless confusion when MATLAB looks like it’s using “half the CPU.” It isn’t — the idle half is virtual. Simultaneous multithreading gives modest system-wide benefit and very little MATLAB-specific gain.

Floating-point units are the hidden bottleneck. On most chips, FPU count matches core count. On some architectures a single FPU is shared between two cores, which quietly throttles numerical work. Nobody advertises this, and it’s part of why two CPUs with identical core counts can behave differently in MATLAB.

Sustained clock speed beats peak clock speed. This is the laptop killer. A thin ultrabook boosts to 5.0 GHz for about eleven seconds, then settles at 2.4 GHz because it physically cannot shed the heat. A thicker chassis with real cooling holds 4.0 GHz all day. For a twenty-minute script, the second machine destroys the first — despite losing on paper.

Buying a laptop for MATLAB? Look up sustained-load benchmarks. A thirty-minute Cinebench loop tells you more than every number on the manufacturer’s page combined.

Two GPU conversations that keep getting mixed up

Drawing things on screen

Plots, 3D surfaces, the Simulink canvas. MathWorks recommends WebGL 2.0 support and at least 2 GB of graphics memory. Integrated graphics — Intel Iris Xe, AMD’s Radeon 700M series, Apple’s M-series GPU — handle this without complaint.

Where a weak GPU shows up: rotating a scatter plot with a million points, or dragging around a sprawling Simulink model. Irritating, not blocking.

On Linux, install the proprietary drivers. The open-source alternatives work in the technical sense and will also make your graphics behave strangely at the worst possible moment.

Doing math on the graphics card

Completely separate topic. This means gpuArray, network training, GPU Coder — real computation offloaded to the GPU. The rules:

• NVIDIA only. Parallel Computing Toolbox works with NVIDIA GPUs at compute capability 3.0 or higher. AMD and Intel GPUs are not supported for computation.
• Not on macOS. GPU acceleration through Parallel Computing Toolbox simply doesn’t exist on Mac. Apple Silicon is superb at plenty of things; CUDA isn’t one of them.
• VRAM outranks raw speed for deep learning. An 8 GB card that fits your model beats a faster 6 GB card that doesn’t. There’s no partial credit — either the model fits or it errors out.

The practical upshot: if your work involves deep learning or GPU Coder, you need NVIDIA hardware and you need Windows or Linux. That’s a wall, not a preference, and it’s the number one reason I’ve watched people return a MacBook.

Simulink rewrites all of the above

Base MATLAB is comparatively light. Simulink is not.

It’s a graphical environment holding entire model hierarchies in memory. Thousands of blocks, referenced models, Simscape physical networks — all of it resident. Add Embedded Coder and you’re now also running a C toolchain that wants its own share.

What I’d budget:

• 16 GB is the floor for serious Simulink work, not the recommendation
• Code generation loves fast NVMe storage because it writes thousands of small files
• Model compilation leans single-threaded, so clock speed matters more here than core count
• Running parsim? Allocate 8 GB per worker and plan accordingly

Platform quirks worth knowing before you commit

Apple Silicon

MATLAB runs natively on M-series chips and runs well. Unified memory suits MATLAB’s contiguous-array habits, and battery performance is in a class of its own.

The catches: no GPU computation, several toolboxes absent, and unified memory means the CPU and GPU share one pool. A 16 GB Mac has less practical headroom than a 16 GB PC with a discrete card, because the display is eating from the same bowl.

My position: 16 GB is the realistic minimum on a Mac and 24–32 GB is what I’d choose. You can’t add more later. That’s a one-time, permanent decision made at checkout.

And if you’re on an Intel Mac — R2025b was the final release. It’ll keep working forever, but you’re frozen there.

Windows on ARM

Snapdragon X machines are supported, but MATLAB runs through Prism emulation rather than natively, and MathWorks notes that certain capabilities aren’t supported on Windows ARM.

For coursework, fine. For anything involving MEX files, hardware support packages, or compiled third-party libraries, buy x86-64 and stop thinking about it.

Linux package gaps

The most common Linux install failure has nothing to do with hardware — it’s missing shared libraries on a stripped-down install. Server images, containers, and minimal desktop setups all hit this.

If MATLAB launches with no window and no error message, that’s your culprit. Make sure the standard desktop package set is present; on Ubuntu, libgtk-3-0, libxt6, libxtst6 and the X11 basics are the usual suspects.

Four builds, actual numbers

Coursework machine — roughly $700–900

Six to eight cores with AVX2, 16 GB RAM, 512 GB NVMe, integrated graphics.

No discrete GPU needed. Put the saved money into RAM and a display you can stare at for eight hours without hating your life.

Research machine — roughly $1,400–1,900

Eight or more cores, 32 GB RAM, 1 TB NVMe, RTX 4060/5060-class GPU with 8 GB VRAM.

The GPU is insurance. The moment your research brushes against machine learning you’ll want it, and you can’t retrofit one into a laptop.

Deep learning workstation — $2,500 and up

Twelve to sixteen cores, 64 GB RAM (128 GB if the budget stretches), 2 TB NVMe, RTX 4080/5080 or better with 16 GB+ VRAM.

Desktop, not laptop. Thermal ceilings make multi-hour training runs miserable on portable hardware, and desktop RAM costs a third as much.

Simulink and embedded work — $2,000 and up

High single-core performance, 32 GB RAM minimum, fast NVMe, Windows.

Favor clock speed over core count here. Compilation and code generation are largely serial, and Windows is where all the toolboxes actually live.

The case for buying nothing

MATLAB Online runs in a browser tab. No installation, no requirements, works on a Chromebook. For coursework-scale work — scripts, plots, modest datasets — it’s a completely legitimate answer that costs nothing extra under most licenses.

It falls apart with large data, long computations, hardware interfacing, and anything needing fast local file access.

There’s also the cluster route. Plenty of universities run MATLAB Parallel Server on shared compute. If you have access, a modest laptop plus batch submission beats an expensive workstation for anything that runs unattended. The sizing guidance there: one worker per physical core, 4 GB RAM per worker, 8 GB per worker if Simulink is involved.

Measure it yourself

MATLAB includes everything needed to answer “is this machine fast for my work” without trusting anyone’s opinion, including mine:

• bench — the built-in benchmark. Good for rough comparison, poor at predicting specific workloads.
• timeit(@() myFunction(args)) — the correct way to time code. Handles warmup, averages multiple runs, produces numbers you can trust. Far better than tic/toc.
• gputimeit — same idea for GPU code.
• memory on Windows — reports available memory and the largest contiguous block. That second figure is what actually predicts out-of-memory failures.

Time the operation you run most often. Whatever it tells you overrules every specification table ever published.

Common questions

Is 8 GB of RAM enough for MATLAB in 2026? For coursework, plotting, and small scripts, yes. For image processing, deep learning, Simulink, or parallel workers, no. MathWorks lists 8 GB as minimum and 16 GB as recommended — treat 16 GB as your real starting point for anything beyond homework.

Do I need a dedicated graphics card? Not for ordinary use. Integrated graphics meeting the WebGL 2.0 recommendation handle plotting fine. A dedicated NVIDIA card is required only for GPU computation: gpuArray, network training, or GPU Coder.

Does MATLAB run on Apple Silicon? Yes, natively, and performance is excellent. Intel Macs are no longer supported as of R2026a — R2025b was their final release. Be aware that GPU computation through Parallel Computing Toolbox isn’t available on macOS at all.

Can I use an AMD or Intel GPU for computation? No. Display and rendering, yes. Computation, NVIDIA only.

How much disk space does MATLAB take? On Windows: 4.6 GB for MATLAB alone, 5–8 GB for a typical install, 25 GB for every product. Mac and Linux are broadly similar. Use an SSD — disk speed is a major factor in start-up time.

Will my older computer eventually stop working? If the CPU lacks AVX2, yes. MathWorks has stated a future release will require it. Roughly 2015 and newer should be safe.

Which operating system is fastest? Performance is broadly comparable. Differences come from compilers, third-party libraries, and how each OS handles disk and graphics operations. Pick based on toolbox availability, not speed.

Is Windows 10 still supported? Yes — version 22H2 specifically, alongside Windows 11 23H2 or newer. Older Windows 10 builds are not.

How much RAM per parallel worker? About 4 GB each, or 8 GB each when Simulink is in the picture. Multiply before you open that pool.

What it comes down to

The official requirements tell you whether MATLAB will start. They say nothing about whether using it will be pleasant.

Compressed into three sentences: get 16 GB of memory as an absolute minimum and 32 GB for anything serious; make sure the processor supports AVX2 so the machine doesn’t age out of compatibility in two years; and install on an SSD, because start-up time is dominated by disk speed. The GPU question answers itself — either your workflow needs CUDA or it doesn’t, and if it does, you need NVIDIA hardware on Windows or Linux.

Everything past that is refinement. Clear those four bars, then stop researching hardware and go write some code.

• Best Laptop for MATLAB
• Best Laptops for Simulink
• What is a good laptop specs for engineering?
• Best Laptops for Java Programming
• Linux 6.15: Powering Up Laptops with Enhanced Hardware Support

Burak Boğa

Hi, my name is Burak. I am a mechanical engineer. I have been writing laptop reviews for the Engineering Laptops website since 2020. Please feel free to contact me if you have any questions.