NPU, GPU and TPU are three kinds of processor that speed up AI, and they are built for different jobs. An NPU (neural processing unit) is a small, low-power accelerator inside a laptop or phone chip, made to run light AI features all day without draining the battery. A GPU (graphics processing unit) is a powerful, flexible parallel processor with its own fast memory, and it is what runs and trains large models in workstations and servers. A TPU (tensor processing unit) is Google's custom AI chip, which you rent through Google Cloud and do not buy. For a business choosing hardware, the practical question is nearly always NPU or GPU, and the answer follows from the size of the model you intend to run.
How do NPU, GPU and TPU compare?
| NPU | GPU | TPU | |
|---|---|---|---|
| What it is | Low-power AI accelerator built into a PC or phone processor | General-purpose parallel processor with dedicated memory | Google's custom AI chip (an ASIC) |
| Built for | Small, always-on inference | Training and inference of any size | Training and inference at data-centre scale |
| Where you find it | AI PCs, phones, some edge devices | Workstations, servers, AI desktops, edge modules | Google Cloud; small Edge TPU modules exist for embedded use |
| Power | A few watts | Tens to hundreds of watts each | Data-centre scale |
| Memory for models | Shares system memory; small models | Dedicated or unified memory, 8GB to well over 100GB | Provided by the cloud service |
| Software support | Varies by vendor; apps must target it | Broadest, especially NVIDIA CUDA | Google frameworks and services |
| How you get it | Comes with the laptop | Buy the card or system | Rent by the hour |
What is an NPU good for?
An NPU handles features that run continuously in the background: blurring a video-call background, cancelling noise, live captions, searching photos, small on-device assistants. It does these using far less power than the main processor or graphics chip would, which is the point. Microsoft's Copilot+ PC label requires an NPU rated at 40 TOPS or more.
What an NPU does not do is run large language models at useful speed, and the TOPS rating says little about that. We examine why in Intel NPU vs NVIDIA GPU: are AI TOPS misleading. A great deal of local AI software also does not use the NPU at all; it targets the GPU.
What is a GPU good for?
Everything larger. A GPU combines thousands of parallel cores with fast memory, and the amount of that memory decides which models fit. It is also where the software is: most AI frameworks, model formats and serving engines are developed and tested on GPUs first. For running a language model for a team, fine-tuning, image and video generation, or computer vision across many cameras, the GPU is the component that matters. See who actually needs 96GB of VRAM and CUDA vs ROCm vs OpenVINO.
What is a TPU, and can you buy one?
A TPU is an application-specific chip designed by Google for the matrix arithmetic of neural networks. It powers Google's own AI services and is offered to customers through Google Cloud. Businesses do not purchase TPUs for their own server rooms; you meet them as a line on a cloud bill. It matters to a buyer mainly as a reminder that the cloud option is not always a GPU, and that work built for one cloud's custom chip is less portable than work built on standard GPUs.
Where do CPUs fit?
The main processor can run small models, slowly, and it runs everything around the model: the operating system, the application, the data handling. Modern AI PCs are designed so that light tasks go to the NPU, heavy tasks to the GPU and the rest to the CPU. Systems with unified memory, where the processor and graphics share one large pool, blur the boundary and allow much larger models than a separate graphics card of the same price. See why 128GB of unified memory matters.
The NPU is for AI features in the software you already use. The GPU is for AI you run yourself. The TPU is someone else's computer.
Which chip does your purchase depend on?
| What you want to do | What to look at |
|---|---|
| Copilot+ features, better video calls, long battery life | NPU: a current AI PC |
| Run small local models on a laptop occasionally | GPU and memory in the laptop; the NPU helps little |
| Private assistant for a team on company documents | GPU with large memory: an AI desktop or workstation |
| Fine-tuning, image or video generation | GPU memory and speed |
| Vision AI on cameras or machines | Edge GPU module such as Jetson |
| Occasional very large jobs, no hardware to manage | Cloud GPU or TPU |
A useful check on any specification sheet: ask how much memory the model will have and which software will actually use the accelerator. Those two answers tell you more than the TOPS number. For a worked comparison of an NPU mini PC against a GPU-class system, see ASUS Ascent QN10 vs GX10, and for edge placement, edge AI vs cloud AI.
Buying in Singapore
AI PCs with an NPU: the GMKtec NucBox K17 (S$1,066.72), GMKtec NucBox K13 (S$1,214.22) and GMKtec EVO-T2S with 64GB (S$3,047.94). GPU-class systems for running models yourself: the NVIDIA DGX Spark 4TB (S$9,450), the ASUS Ascent GX10 1TB (S$10,350) and the NVIDIA RTX PRO 6000 Blackwell Workstation Edition with 96GB (S$29,100). At the edge, the NVIDIA Jetson Orin Nano Super Developer Kit (S$696) (listed 11 Nov 2026; prices change; ask for a current quotation). To start, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
What is the difference between an NPU and a GPU? An NPU is a small, low-power accelerator built into a PC or phone chip for light, always-on AI features. A GPU is a much more powerful processor with its own memory that runs and trains large AI models.
What is a TPU? A TPU, or tensor processing unit, is a custom AI chip designed by Google. It is used in Google's data centres and rented through Google Cloud; businesses do not buy it as hardware.
Can an NPU run a large language model? Not at useful speed for large models. NPUs suit small models and background features. Larger language models need a GPU with enough memory.
Do I need an NPU in a business laptop? It is worthwhile for Copilot+ features, video-call effects and battery life, and most new business laptops include one. It is not a substitute for a GPU if you plan to run models yourself.
