Laptop Buying Advice

A practical guide for students of the Department of Advanced Computing Sciences at Maastricht University, especially:

BSc Computer Science
BSc Data Science and Artificial Intelligence
MSc Artificial Intelligence
MSc Data Science for Decision Making
MSc Responsible Data Science

Who is the AI hardware advice for? The sections about CUDA, GPU memory, local model inference, and neural-network training are aimed mainly at BSc Data Science and Artificial Intelligence students, MSc Artificial Intelligence students, and anyone intentionally choosing AI- or deep-learning-heavy electives or research projects. They are not general hardware requirements for every DACS student. In particular, most Computer Science, Data Science for Decision Making, and Responsible Data Science coursework does not require an AI-focused laptop.

Short recommendation

For a new laptop, we recommend at least:

16 GB RAM (32 GB is preferable for serious local AI, large datasets, virtual machines, or a laptop you intend to keep for several years)
512 GB SSD storage (1 TB is preferable for datasets, virtual machines, containers, and local AI models)
A recent mid-range or better processor, such as an Intel Core Ultra 5/Core i5, AMD Ryzen 5, Apple M-series chip, or equivalent
A recent version of Windows, macOS, or Linux
A good keyboard, screen, battery, and build quality: you will use this machine every day

For most coursework, 16 GB RAM and 512 GB storage are sufficient. You do not need an expensive gaming laptop merely to complete your degree. University servers and cloud resources may be used for workloads that are too large for a laptop.

If you specifically want to train neural networks locally or work with software that expects CUDA, choose a laptop with an NVIDIA RTX GPU. Check its VRAM, not only the GPU model name.

Do you need a laptop?

Yes. A laptop is used in computer labs, group projects, assignments, and presentations. It should be able to run development tools for languages such as Python, Java, C/C++, and R, as well as browsers, office software, containers, and possibly virtual machines.

The programmes do not all have the same emphasis. See Suggested configurations by programme for specific advice.

Course requirements can change. If a particular course publishes hardware or operating-system requirements, follow those over this general guide.

Memory and storage

RAM

16 GB is the minimum we recommend for a new laptop in 2026. Avoid buying an 8 GB machine unless it is an inexpensive temporary solution and its memory can be upgraded. Many thin laptops have soldered memory, so check before buying.

Choose 32 GB if you expect to use several of the following at once:

IDEs, notebooks, and many browser tabs
Docker containers or virtual machines
Large datasets or in-memory data processing
Local language, vision, or generative models
GPU workloads that can offload data to system memory

For unusually large local workloads, 64 GB may be useful, but it is not a general degree requirement.

SSD storage

512 GB is the minimum we recommend. Operating systems, development environments, Docker images, datasets, virtual machines, and model weights consume space quickly. Choose 1 TB if the upgrade is affordable, especially when storage cannot be replaced later.

Prefer an SSD over a hard drive. External storage is useful for archives and backups, but it is less convenient for active projects and datasets.

Understanding GPUs for AI and machine learning

This section is primarily relevant to DSAI and AI-focused students. Other students may safely treat it as optional buying advice unless their electives, thesis, research, or personal projects involve machine learning.

The GPU is the most confusing part of buying an AI laptop. A GPU can make some machine-learning workloads much faster, but many introductory algorithms and data-science tasks still run on the CPU. A fast GPU also cannot compensate for insufficient memory.

NVIDIA GPUs: the safest option for local training

NVIDIA remains the most compatible choice for deep learning because CUDA is widely assumed by research code, libraries, tutorials, and course projects. If local training matters to you, look for a recent GeForce RTX laptop GPU.

Pay close attention to VRAM:

6 GB: usable for learning and small models, but restrictive
8 GB: a reasonable entry point for smaller deep-learning experiments
12 GB or more: preferable for local training, larger models, and more flexibility
16 GB or more: excellent for a laptop, but expensive and still not enough for every workload

Laptop GPUs with the same product family name can have different power limits and performance. Read independent reviews of the exact laptop, not just the GPU specification.

Apple GPUs: strong inference, different training ecosystem

Apple silicon MacBooks are fast, quiet, and power-efficient. Their CPU and GPU share a pool of unified memory, which can make a Mac with sufficient memory particularly good for running quantised local language models and other inference workloads. There is no separate VRAM capacity in the same sense as on a discrete NVIDIA laptop GPU.

Apple GPUs do not support CUDA. PyTorch can accelerate both inference and training through the MPS/Metal backend, and many popular local-inference applications support Apple silicon. However, some operations, libraries, research repositories, and course instructions are written specifically for CUDA. They may need workarounds, fall back to the CPU, or not work at all.

In short: a MacBook is a good general development laptop and can be excellent for local inference. It is less convenient for CUDA-based training and reproducibility. If buying one for AI work, prefer at least 16 GB unified memory, and seriously consider 24 GB or more. Remember that macOS and applications also use this same memory pool.

AMD GPUs

AMD GPUs can accelerate machine learning through ROCm and other backends. Support has improved substantially, including support for selected Radeon GPUs and Ryzen AI processors, but it remains more dependent on the exact GPU, operating system, framework, and software version than CUDA.

AMD can be a good option for an informed buyer who has checked the current compatibility matrix for the exact model. It is not the safest default when you need arbitrary research code to work without modification. Integrated Radeon graphics share system RAM and are mainly suited to light acceleration and inference, not heavy training.

Intel GPUs

Recent Intel Arc discrete GPUs and Intel integrated GPUs can accelerate supported AI workloads through technologies such as the PyTorch XPU backend, oneAPI, OpenVINO, and platform-specific inference runtimes. They are increasingly useful for inference and experimentation.

As with AMD, compatibility depends on the exact hardware and software stack. Intel GPUs do not provide CUDA, and code written only for CUDA may require changes. An Intel Arc GPU can be a reasonable value choice for supported workloads, but NVIDIA remains the lower-friction option for general local deep-learning training.

Integrated GPUs and NPUs

Integrated GPUs from Intel and AMD share system RAM with the CPU. They are useful for display work, media, light GPU computing, and some optimised inference workloads. Newer integrated GPUs are much more capable than older ones, but they usually have lower sustained performance and memory bandwidth than a good discrete GPU.

Many new processors also include an NPU. NPUs are efficient for supported on-device inference, such as transcription, image processing, or operating-system AI features. They are not a replacement for a CUDA GPU for general model training, and framework support varies.

An integrated GPU or NPU is perfectly adequate if you plan to do heavy computation on university servers or in the cloud. Do not pay a large premium for an “AI PC” label without checking which tools can actually use its NPU.

VRAM, shared memory, and model size

For AI workloads, available memory often matters more than peak GPU speed. Model weights, activations, gradients, optimiser state, and data batches all consume memory. Training normally needs far more memory than inference.

On a laptop with a discrete GPU, VRAM and system RAM are separate. Some frameworks can offload model layers or data from VRAM into system RAM when the model does not fit. This is not guaranteed to happen automatically, and it is much slower because data must move between the CPU and GPU. If you intend to rely on offloading, buy enough system RAM as well: 32 GB or more is sensible.

Apple silicon and some integrated-GPU systems use shared or unified memory. This can allow the GPU to access a larger memory pool, but the operating system and applications need part of that pool too. Large memory capacity makes a model possible to run; it does not necessarily make it fast.

Quantisation can greatly reduce the memory required for inference. Large-scale training, full fine-tuning, and high-throughput experiments generally belong on a server, desktop workstation, or cloud platform rather than a student laptop.

Chromebooks, tablets, and unusual devices

We do not recommend a Chromebook as your primary study laptop. Some Chromebooks provide a Linux development environment and can run editors, terminals, and lightweight programming tools, but compatibility with virtualisation, containers, specialised drivers, required desktop software, and GPU computing is inconsistent. Storage and RAM are often limited as well.

An Android tablet or iPad is useful as a companion for notes and reading, but it should not be your only computer. Windows-on-ARM laptops may be excellent portable machines, but check that required development tools, virtualisation software, and drivers support ARM before buying.

If you already own one of these devices, it may cover part of the programme with remote access to another machine. That is different from recommending it as a safe new purchase.

Operating system

Windows: broadly compatible and supports NVIDIA CUDA. WSL 2 provides a convenient Linux environment, although virtualisation and storage require extra disk space.
Linux: excellent for development and closely matches many servers. Hardware support, battery life, fingerprint readers, sleep, and vendor utilities vary by laptop model.
macOS: excellent Unix-based development environment and strong Apple-silicon efficiency. It does not support CUDA, and some x86-only or Windows-only tools may require alternatives or virtualisation.

No operating system is universally best. Choose one you can work with, while considering the GPU software you expect to use.

Other things worth checking

Screen: 14 to 16 inches is a practical range. Prefer a readable resolution and enough brightness over an unnecessarily high refresh rate.
Keyboard layout: make sure it is one you are comfortable using, such as QWERTY, QWERTZ, or AZERTY.
Battery and weight: a powerful gaming laptop may spend much of the day near a power socket and can be tiring to carry.
Cooling: sustained CPU and GPU performance depends on cooling. Thin designs may slow down under long workloads.
Ports and charging: check for the displays, USB devices, and chargers you use.
Repairability: replaceable storage, memory, and battery can extend the useful life of a laptop.
Warranty: international students should check where warranty service is available.

Suggested configurations by programme

These specifications are intended for a new primary laptop. “Minimum” means a practical floor for completing the programme, not the cheapest computer that could technically run an editor. “Recommended” provides more room for several years of coursework, development tools, datasets, containers, and virtual machines.

Programme	Minimum	Recommended	GPU advice
BSc Computer Science	16 GB RAM, 512 GB SSD	32 GB RAM, 1 TB SSD	A dedicated GPU is recommended, especially for graphics, game development, computer vision, parallel computing, and AI electives. It is not required for most core programming work.
BSc Data Science and Artificial Intelligence	16 GB RAM, 512 GB SSD	32 GB RAM, 1 TB SSD	For local deep learning, choose an NVIDIA GPU with at least 8 GB VRAM. 12 GB or more provides substantially more flexibility.
MSc Artificial Intelligence	16 GB RAM, 512 GB SSD	32 GB RAM, 1 TB SSD	An NVIDIA GPU with 8 GB VRAM is the practical entry point for local AI work; 12 GB or more is strongly preferred. Large training jobs will still require university or cloud compute.
MSc Data Science for Decision Making	16 GB RAM, 512 GB SSD	32 GB RAM, 1 TB SSD	Prioritise RAM and CPU performance for data processing, optimisation, notebooks, and containers. A dedicated GPU is optional unless electives, thesis work, or other projects involve deep learning.
MSc Responsible Data Science	16 GB RAM, 512 GB SSD	32 GB RAM, 1 TB SSD if the budget allows	Prioritise RAM, CPU performance, portability, battery life, and software compatibility. A dedicated GPU depends on electives, thesis work, and research direction.

Choosing the CPU

Choose a recent mid-range or better processor, such as an Intel Core Ultra 5/Core i5, AMD Ryzen 5, Apple M-series chip, or equivalent. Avoid Intel N-series, Celeron, Pentium, and similarly low-end processors in a new primary laptop.

A Core Ultra 7/Core i7, Ryzen 7, or higher-tier Apple chip can help with compilation, data processing, and CPU-based modelling, but do not overpay for the CPU if it forces you to accept too little RAM, storage, or GPU memory. Cooling also matters: a high-end processor in a very thin laptop may not sustain its advertised performance.

Choosing a dedicated GPU

For BSc Computer Science, “dedicated GPU recommended” does not mean that any dedicated GPU is worth buying. Do not sacrifice RAM, storage, battery life, or build quality merely to obtain a weak GPU. If AI or GPU computing is the reason for buying one, an NVIDIA RTX GPU is generally preferable because of CUDA support.

For local AI work, use the following VRAM guidance:

Less than 8 GB: suitable for light acceleration and small experiments, but difficult to recommend for a new AI-focused laptop
8 GB: practical entry point for smaller training jobs, computer vision, and modest local inference
12 GB: preferred starting point for meaningful local AI experimentation
16 GB or more: desirable for larger models and workloads, although it still does not replace server hardware

A faster GPU with insufficient VRAM may be less useful for AI than a slower GPU with more VRAM. Always verify the VRAM and power limit of the exact laptop GPU; the product name alone is not enough.

Apple configuration for local inference

For a MacBook, consider the unified-memory capacity separately from the minimum needed for coursework:

16 GB: coursework, development, and light local inference
24-32 GB: a practical range for local inference and moderate models
48-64 GB or more: useful for larger quantised models that cannot fit in typical laptop VRAM

The operating system and applications share this memory with the GPU, so not all of it is available to the model. More unified memory can make a model possible to run, but does not make macOS compatible with CUDA-based software.

Budget priority

If the budget cannot cover every upgrade, use this general order:

Meet the 16 GB RAM and 512 GB SSD minimum.
Increase RAM to 32 GB, particularly for DSAI and MSc programmes.
Increase storage to 1 TB.
Add an appropriate dedicated GPU if your intended work benefits from it.

For an AI-focused purchase, GPU memory can move ahead of SSD capacity once the minimum storage requirement is met. Avoid buying a powerful GPU in a system with inadequate RAM, storage, cooling, or battery life.

Final advice

Do not choose a laptop from CPU or GPU branding alone. Check the exact amount of RAM, SSD storage, GPU memory, power limits, upgradeability, battery life, and independent reviews of the complete model.

For most students, 16 GB RAM, a 512 GB SSD, and a recent mid-range processor are the floor. If your budget allows one major upgrade, 32 GB RAM or a 1 TB SSD will often improve everyday student life more than a low-end dedicated GPU. Choose NVIDIA when local CUDA work is a priority; choose Apple silicon when portability, battery life, and local inference matter more than CUDA compatibility.