What is the difference between a GPU, a TPU and an NPU?

An AI model is mostly a very long list of numbers. To answer you, it multiplies those numbers against your input, adds up the results, and repeats, layer after layer, for every word it writes. The main processor in your computer, the CPU, is built to do a few varied tasks quickly in sequence. AI work calls for the opposite: one simple task repeated billions of times, side by side.

That is the job all three chips share. They differ in who makes them, how specialized they are, and how much power and memory they have to work with.

The three chips in one table

	GPU	TPU	NPU
Stands for	Graphics processing unit	Tensor Processing Unit	Neural processing unit
Who makes it	Nvidia, AMD and others	Google	Apple, Qualcomm, Intel, AMD and others, inside their processors
Where it runs	Gaming PCs, workstations, data centres	Google data centres	Phones and laptops
Best at	Almost any parallel work, including training and serving large models	Training and serving large models at Google scale	Small models on battery power
How you get one	Buy it, or rent it from a cloud	Rent it through Google Cloud	It comes with your device

"Tensor" is the word for the grids of numbers these models work on, which is where the TPU gets its name.

GPU: the flexible one

GPUs began as graphics chips, because drawing a video game frame also means doing the same small calculation on millions of pixels at once. That design suits neural networks, and it stays programmable, so the same chip can run new kinds of models as they appear.

At the top end, a GPU is no longer a single card you hold in your hand. Nvidia's Vera Rubin NVL72 is a whole rack that links 72 Rubin GPUs and 36 Vera CPUs so they can work as one machine. Nvidia's specification for a 100-megawatt site lists 40,000 Rubin GPUs and 12 petabytes of fast memory, which works out to about 300 GB per GPU. Those figures come from the vendor, and its performance claims are marked as projections.

TPU: the specialist

Google began running its first TPU in its own data centres in 2015. The paper describing it shows how narrow the design was: one large unit of 65,536 multipliers working on 8-bit numbers, reaching a peak of 92 trillion operations per second, used only to serve models that had already been trained.

Removing what a neural network does not need saves energy and cost. The trade-off is less flexibility and a software stack that has mostly grown around Google's own tools.

The line has since split in two. In April 2026, Google announced its eighth generation as two chips, and said they would reach general availability later in the year:

Chip	Built for	Figures Google gives
TPU 8t	Training	Pods of up to 9,600 chips with 2 petabytes of shared fast memory
TPU 8i	Inference (answering requests)	288 GB of fast memory and 384 MB of on-chip memory per chip

Google says both can run either kind of work, and that specializing simply makes each more efficient. TPUs are also used outside Google: Anthropic says it trains and runs Claude on a mix of AWS Trainium chips, Google TPUs and Nvidia GPUs. Trainium is Amazon's in-house chip, in the same family of purpose-built designs as the TPU.

NPU: the frugal one

An NPU is a section of the processor in your phone or laptop, designed to run AI features without draining the battery. Blurring a video background, transcribing a call or running a small assistant are typical jobs.

Its speed is usually quoted in TOPS, trillions of operations per second. Microsoft's bar for a Copilot+ PC is more than 40 TOPS. That is under half the 92 TOPS of Google's 2015 chip, though the two figures were not measured on the same work, so treat the comparison as a sense of scale only.

The categories also overlap more each year. Apple's M5 has a separate Neural Engine, and also puts a Neural Accelerator inside each of the 10 GPU cores. On that chip, a local model may well run on the GPU rather than the NPU.

A worked example: why memory decides

Speed is only half of it. A model must fit in memory before it can run at all. Take Kimi K2, an open model with 1 trillion parameters. At one byte per parameter, a common compressed format, that is about 1 TB of memory.

Hardware	Memory	Does Kimi K2 fit?
Laptop with 16 GB	16 GB shared	No, about 60 times too small
TPU 8i	288 GB per chip	Needs at least 4 chips
Rubin GPU	About 300 GB per chip	Needs at least 4 chips

The arithmetic is simplified and ignores working memory, but it shows why data-centre chips ship with so much fast memory, and why that memory is itself in short supply.

Try it: Open your computer's system information and find its processor name. Search for that name plus "NPU TOPS" and "GPU memory", then note the two numbers. They are the starting point for deciding what will run on your machine.

Which of these chips do you actually need?

If you only use AI through a website or an app, none of them. The chips sit in someone else's data centre, and your device just sends text back and forth. An NPU may make a few built-in features faster or more private, but it will not change what ChatGPT, Claude or Gemini can do for you.

If you want to run models yourself, memory comes first. A GPU with plenty of video memory, or a Mac with plenty of unified memory, gives you the widest choice of models and the best-supported tools. An NPU in a laptop is mainly useful for small models and the features your operating system or apps already build on it.

If you are training or serving models for a company, you will mostly rent. Nvidia GPUs have the broadest software support. TPUs are worth pricing if your work already fits Google Cloud and its tools. The right test is your own workload: run the same job on each option you are considering and compare the cost per result.