ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›BLOG›GPU VS NPU VS TPU: WHAT IS THE DIFFERENCE?
BLOG · COMPARISONBEGINNER

GPU vs NPU vs TPU: What Is the Difference?

A GPU is a powerful chip for big AI work, an NPU a small low-power AI chip in phones and laptops, and a TPU Google’s own AI chip, used through Google Cloud.

READ8 min
WORDS1,504
SECTIONS8
TYPECOMPARISON
CHECKED19 OCT 26

A GPU is a powerful chip first made for graphics that now trains and runs many large AI models. An NPU is a small, low-power AI chip inside phones and newer laptops. A TPU is Google’s own custom AI chip, used in its data centres and rented out through Google Cloud. All three do many small sums at once; they differ in size, power use and where you find them.

TL;DR — THE SHORT VERSION
  • GPU (graphics processing unit): the big, flexible workhorse. In gaming PCs and in data centres.
  • NPU (neural processing unit): the small, battery-friendly AI chip in your phone or laptop.
  • TPU (tensor processing unit): Google’s own AI chip, built for the grid maths at the heart of AI. You reach it through Google Cloud.
  • For using a chatbot, you need none of them. The company’s chips do the work.
◈ IN PLAIN TERMS

Think of kitchens. The ordinary processor (CPU) is a chef who can cook anything, one dish at a time. A GPU is a huge, power-hungry oven that bakes thousands of trays at once. An NPU is a kettle: small, efficient, made for one quick job at home. A TPU is a factory bakery one company built for its own recipes, where others can rent time on the ovens.

GPU, NPU and TPU at a glance
ChipWhat it isWhere you find itGood for
GPUIs: a chip with thousands of small cores, first built for graphics.Found: graphics cards in PCs; racks in data centres.Good for: training and running large models; running AI at home.
NPUIs: a small chip made only for AI, built to save power.Found: inside phones and newer laptops, on the same chip as the other processors.Good for: quick AI features on the device, such as live captions and video-call effects.
TPUIs: Google’s custom chip, designed to do matrix maths quickly.Found: Google’s data centres; used through Google Cloud services.Good for: large models dominated by matrix maths that train for weeks or months.

What do GPUs, NPUs and TPUs have in common?

The plain point: all three are built to do huge numbers of simple sums at the same time, which is what AI needs.

An AI model’s work is mostly grids of numbers, called matrices, multiplied together. That job splits neatly into thousands of small pieces. IBM explains that a GPU’s strength is “hundreds—even thousands—of cores, enabling the parallel processing” that makes it fast.1 It says NPUs “can perform similar parallel processing with far better power efficiency.”2 And Google says its TPUs “are designed to perform matrix operations quickly making them ideal for machine learning workloads.”3

The ordinary processor, the CPU, can do the same maths, just more slowly: IBM notes that CPUs “use fewer cores and perform their tasks in a linear order.”4

What is a GPU?

The plain point: a graphics chip that turned out to be very good at AI maths.

IBM describes the GPU as a circuit designed “to speed computer graphics and image processing”. Its guide dates the first GPU chip, Nvidia’s GeForce, to 1999, and says Nvidia’s CUDA software, built in 2007, gave programmers direct access to the chip’s parallel computing for work beyond graphics.5 Today, IBM notes, “AI models train and run on data center GPUs”.6

A gaming PC has one too, on its graphics card. If you want to run a model on your own computer, the GPU and its memory, called VRAM, are usually the limit. More on GPU meaning.

What is an NPU?

The plain point: a small AI specialist that lives inside your device and saves battery.

IBM says NPUs, unlike general-purpose CPUs or GPUs, “are tailored to accelerate AI tasks and workloads”, and that phones and laptops usually combine the NPU with other processors on one chip, a “system-on-chip (SoC)”.7 Microsoft builds its Copilot+ PCs around an NPU, which it says lets “AI tasks like real-time translation, image generation, and intelligent search to run locally.”8

An NPU handles small, quick jobs. The large models behind the leading chatbots still run on big chips in data centres. More on NPU meaning.

What is a TPU?

The plain point: Google’s own AI chip, running in Google’s data centres. Other developers rent time on it through Google Cloud.

Google’s documentation calls TPUs “Google’s custom-developed, application-specific integrated circuits (ASICs)” used to speed up machine learning. An ASIC is a chip made for one kind of job.9 Google announced the chip in 2016, saying it had already been “running TPUs inside our data centers for more than a year”, and that TPUs powered AlphaGo in its matches against the Go champion Lee Sedol.10 Google also says it trained its Gemini 1.0 models on TPUs.11

Outside Google, developers use them through Google Cloud: “You can use TPUs through Compute Engine, Google Kubernetes Engine, and Gemini Enterprise Agent Platform.”9 Google’s architecture page listed seven TPU versions on 19 Oct 2026, from TPU v2 to TPU7x, also called Ironwood.3

When is each chip used?

The plain point: even Google, which makes TPUs, says each chip has jobs it suits and jobs it does not.

Google’s own guide suggests GPUs for “Medium-to-large models with larger effective batch sizes” and for models with many custom operations, and TPUs for “Models dominated by matrix computations”, “Models that train for weeks or months” and “Large models with large effective batch sizes”. It lists work TPUs are not suited for, including “Workloads that require high-precision arithmetic”.12 NPUs sit at the other end of the scale: small jobs on your own device, where saving battery matters most.

TAKEAWAY

GPU for big, flexible AI work and for running models at home; NPU for small AI features on your phone or laptop; TPU for large training and serving jobs on Google’s cloud.

Which one do I need?

The plain point: for most people, the answer is none. This section is this site’s reading of the sources above.

  • You use ChatGPT, Claude or Gemini in a browser: the work happens on the company’s chips. Your own hardware hardly matters.
  • You want to run a model on your own computer: look at your GPU and its VRAM first. Run AI on your own computer shows how to check.
  • You are choosing a laptop for AI features: an NPU helps with built-in features, but check that the apps you use can use it.
  • You are a developer training a large model: GPUs, and TPUs on Google Cloud, are both options; Google’s guide above helps decide.

Questions people ask

Is a TPU better than a GPU?

Not across the board. Google’s own guide recommends TPUs for some jobs and GPUs for others, and lists work TPUs are not suited for (source above).

What is the difference between an NPU and a TPU?

Both are chips made for AI. An NPU is small and sits inside your phone or laptop; a TPU is Google’s data-centre chip that you reach through Google Cloud (sources above).

Does my phone have an NPU?

Possibly: IBM says smartphones and laptops usually combine the NPU with other processors on one chip (source above). Check your phone maker’s specification page to be sure.

Do I need a GPU to use AI?

Not to use a chatbot: the heavy work runs on the company’s computers. You only need a good GPU to run models yourself; see local AI meaning.

Where to go next

QUICK CHECK · 3 QUESTIONS

Pick an answer to see if you got it. Nothing is recorded.

1. What do GPUs, NPUs and TPUs have in common?

  • They all do many small sums at the same time
  • They are all made by Google
  • They are all inside every phone
Show the answer

Answer: They all do many small sums at the same time. The page says all three are built to do huge numbers of simple sums at once, which is what AI maths needs.

2. Where would you usually find an NPU?

  • Only in Google’s data centres
  • Inside phones and newer laptops
  • On a graphics card in a gaming PC
Show the answer

Answer: Inside phones and newer laptops. The page describes the NPU as a small, low-power AI chip inside phones and newer laptops.

3. How do developers outside Google use TPUs?

  • They plug one into their laptop
  • They download it from a website
  • Through Google Cloud services
Show the answer

Answer: Through Google Cloud services. The page quotes Google: TPUs are used through Compute Engine, Google Kubernetes Engine and Gemini Enterprise Agent Platform.

SOURCES
  1. 1IBM, What is a graphics processing unit (GPU)?, undated, read at source 1 Oct 2026 (as cited on this site’s GPU page): “GPUs, however, have hundreds—even thousands—of cores, enabling the parallel processing that drives their lightning-fast processing capabilities.”↑
  2. 2IBM, What is a Neural Processing Unit (NPU)?, undated, read at source 1 Oct 2026 (as cited on this site’s NPU page): “NPUs can perform similar parallel processing with far better power efficiency.”↑
  3. 3Google Cloud, TPU architecture, last updated 13 Oct 2026, read at source 19 Oct 2026“TPUs are designed to perform matrix operations quickly making them ideal for machine learning workloads.” ↑aversions listed TPU7x (Ironwood), TPU v6e, TPU v5p, TPU v5e, TPU v4, TPU v3 and TPU v2. ↑b
  4. 4IBM, What is a graphics processing unit (GPU)?, undated, read at source 12 Oct 2026 (as cited on this site’s GPU page): “CPUs tend to use fewer cores and perform their tasks in a linear order.”↑
  5. 5IBM, What is a graphics processing unit (GPU)?, undated, read at source 12 Oct 2026 (as cited on this site’s GPU page): “The first GPU chip, the GeForce from Nvidia, was released in 1999” and “In 2007, Nvidia built CUDA™ (Compute Unified Device Architecture), a software platform and” programming interface giving developers direct access to the GPU’s parallel computing.↑
  6. 6IBM, What is a graphics processing unit (GPU)?, undated, read at source 12 Oct 2026 (as cited on this site’s GPU page): “AI models train and run on data center GPUs, typically operated by enterprises conducting scientific research”.↑
  7. 7IBM, What is a Neural Processing Unit (NPU)?, undated, read at source 1 Oct 2026 (as cited on this site’s NPU page): “Differing from general-purpose central processing units (CPUs) or graphics processing units (GPUs), NPUs are tailored to accelerate AI tasks and workloads”; “most consumer applications, such as smartphones, mobile devices and laptops, combine the NPU with other coprocessors on a single semiconductor microchip known as a system-on-chip (SoC).”↑
  8. 8Microsoft, Copilot+ PCs & Windows PCs: Differences?, 19 Dec 2025, read at source 1 Oct 2026 (as cited on this site’s NPU page): “enabling AI tasks like real-time translation, image generation, and intelligent search to run locally.”↑
  9. 9Google Cloud, Introduction to Cloud TPU, last updated 13 Oct 2026, read at source 19 Oct 2026“Tensor Processing Units (TPUs) are Google’s custom-developed, application-specific integrated circuits (ASICs) used to accelerate machine learning workloads.” ↑a“You can use TPUs through Compute Engine, Google Kubernetes Engine, and Gemini Enterprise Agent Platform.” ↑b
  10. 10Google Cloud blog, Google supercharges machine learning tasks with TPU custom chip, dated 19 May 2016 on the page, read at source 19 Oct 2026“We’ve been running TPUs inside our data centers for more than a year”; “AlphaGo was powered by TPUs in the matches against Go world champion, Lee Sedol”. ↑
  11. 11Google, Introducing Gemini: our largest and most capable AI model, 6 Dec 2023, read at source 19 Oct 2026“We trained Gemini 1.0 at scale on our AI-optimized infrastructure using Google’s in-house designed Tensor Processing Units” (v4 and v5e). ↑
  12. 12Google Cloud, Introduction to Cloud TPU, last updated 13 Oct 2026, read at source 19 Oct 2026, “When to use TPUs”GPUs, “Medium-to-large models with larger effective batch sizes”; TPUs, “Models dominated by matrix computations”, “Models that train for weeks or months”, “Large models with large effective batch sizes”; not suited, “Workloads that require high-precision arithmetic”. ↑

Read at source 19 Oct 2026: Google Cloud, Introduction to Cloud TPU and TPU architecture (both last updated 13 Oct 2026); Google Cloud blog, Google supercharges machine learning tasks with TPU custom chip (19 May 2016); Google, Introducing Gemini (6 Dec 2023).

Reused from this site’s pages, with the read dates recorded there: IBM’s GPU guide (1 and 12 Oct 2026), on GPU meaning; IBM’s NPU guide and Microsoft’s Copilot+ PC page (1 Oct 2026), on NPU meaning.

The kitchen comparison and the “which one do I need” advice are this site’s own ways of explaining, not claims from these sources.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE