Edited by 
Short answer
An NPU (neural processing unit) is a small AI accelerator built into phone, laptop and PC processors. It runs on-device AI such as background blur, live captions and small language models using a few watts. A GPU is a large, programmable parallel processor; with Tensor Cores it runs anything from games to training giant models, at hundreds of watts. A TPU is Google's data-center AI chip, built around huge matrix units and connected by the thousand into pods. All three speed up the same core operation, matrix multiplication, at very different scales: about 40 to 80 TOPS for a laptop NPU, about 3,352 AI TOPS for an RTX 5090, and pods of thousands of chips for TPUs.
NPU vs GPU vs TPU at a glance
| NPU | GPU | TPU | |
|---|---|---|---|
| Where you find it | Inside phone, laptop and PC processors | Graphics cards, workstations, data centers | Google data centers and Google Cloud |
| Main job | Efficient on-device inference | Graphics plus training and inference of any size | Training and serving very large models |
| Typical power | A few watts | 72 W to 1,400 W | Not published, data-center class |
| Typical rating | 40–80 TOPS, INT8 | Hundreds to thousands of TFLOPS | About 4,600 FP8 TFLOPS (Ironwood) |
| Memory | Shares system RAM | Own GDDR or HBM | Own HBM (192 GB on Ironwood) |
| Programmability | Limited, through vendor runtimes | Fully programmable (CUDA, ROCm) | Through XLA, JAX and PyTorch |
| Can you buy one? | Yes, in a new laptop or phone | Yes | No, rent it in Google Cloud |
What is an NPU?
An NPU is a block of silicon dedicated to neural-network math, placed next to the CPU and GPU on the same chip. Like a Tensor Core, its heart is an array of multiply-accumulate units that process small matrices in one step. Unlike a GPU, it gives up most of its flexibility to save power: it supports fewer operations and formats, mostly INT8 and FP16, and it is reached through vendor runtimes such as Windows ML, Core ML or Qualcomm's AI Engine rather than general-purpose code.
That trade-off makes sense for AI that runs all day in the background. Blurring your webcam background on the GPU would drain a laptop battery quickly; on the NPU it costs a fraction of a watt.
NPUs you can buy in 2026
| Processor | NPU | NPU rating |
|---|---|---|
| Qualcomm Snapdragon X2 Elite | Hexagon | 80 TOPS |
| Intel Core Ultra series 3 (Panther Lake) | NPU 5 | 50 TOPS |
| AMD Ryzen AI 300 | XDNA 2 | 50 TOPS |
| Intel Core Ultra 200V (Lunar Lake) | NPU 4 | 48 TOPS |
| Qualcomm Snapdragon X Elite | Hexagon | 45 TOPS |
| Apple M4 | 16-core Neural Engine | 38 TOPS |
Microsoft requires an NPU of at least 40 TOPS for a laptop to carry the Copilot+ PC label. Apple has not published a TOPS figure for the M5's Neural Engine; on M5 it also added neural accelerators to each GPU core.
How a GPU differs
A GPU is built from many streaming multiprocessors, each with general-purpose cores and Tensor Cores (see how a GPU works). It is fully programmable, supports every common number format from FP64 to FP4, and has its own very fast memory: 1.79 TB/s of GDDR7 on an RTX 5090 and 8 TB/s of HBM3e on a B200. That is why GPUs, not NPUs, run large language models at useful speed. An NPU shares ordinary laptop memory with the CPU, typically around 100 to 150 GB/s, and generating each token has to read the model's weights from that memory.
How a TPU differs
A TPU is an ASIC that Google designs for its own data centers. Each chip has large systolic arrays, grids of multiply-accumulate cells that pass data between neighbors, plus HBM and a dedicated chip-to-chip network. Google's 2025 Ironwood TPU has 192 GB of HBM3e at 7.4 TB/s and 4,614 dense FP8 TFLOPS per chip, in pods of up to 9,216 chips. In spirit a TPU is a giant data-center NPU; in scale it competes with NVIDIA's largest GPUs. Read TPU vs GPU for the full comparison.
Why the TOPS numbers don't compare directly
An NPU's 50 TOPS and an RTX 5090's 3,352 AI TOPS look like a 67-times gap, but they measure different things. NPU figures are usually INT8 without sparsity. NVIDIA's AI TOPS for GeForce cards are FP4 with 2:4 sparsity, which counts each operation four times as generously as dense INT8 would. On an equal basis the gap is smaller, though still very large. Our guide to TOPS vs TFLOPS shows how to convert between them.
Which one runs your AI?
- Background and system features such as noise suppression, live captions, image search and on-device assistants: the NPU, to save battery.
- Running a local LLM or image model: the GPU, or Apple silicon's unified memory, because speed depends on memory bandwidth. See the best GPU for local LLMs.
- Training and serving models at scale: data-center GPUs or TPUs, rented in the cloud.
Frequently asked questions
Do I need an NPU to run AI on my laptop?
No. Most local AI apps, including tools for running language models such as Ollama and LM Studio, use the GPU or CPU. An NPU mainly helps battery life for always-on features and is required for some Windows Copilot+ features.
Can an NPU run ChatGPT or a large language model?
ChatGPT itself runs in OpenAI's data centers. NPUs can run small language models of a few billion parameters, such as those built into Windows and phones. Larger models are limited by the memory bandwidth the NPU shares with the rest of the chip, so a GPU is much faster for them.
Is a Tensor Core an NPU?
No. A Tensor Core is a matrix unit inside each streaming multiprocessor of an NVIDIA GPU, working alongside the GPU's other cores. An NPU is a separate accelerator block on a processor. Both are built to do matrix multiply-accumulate quickly.
Is a TPU better than a GPU?
Neither is better in general. TPUs are efficient for large models written in JAX or run through XLA on Google Cloud. GPUs run nearly every framework and model and are available from every cloud and for purchase.
