Learn AI /Data and Infrastructure /GPU, CPU and TPU
GPU, CPU and TPU
- Starter
- Data and Infrastructure
- Policy officers
- Managers & front offices
Also called: GPU, CPU, TPU, chips, accelerator.
Back to the milestone you came from
In five seconds
What it is
CPUs, GPUs and TPUs are different kinds of computer chips that run calculations, with each designed for different patterns of work, speed, cost and energy use.
In plain words
Think of CPUs, GPUs and TPUs as different kinds of engines for digital work. CPUs are flexible all-rounders. GPUs are good at doing many similar calculations at once. TPUs are specialised chips for certain AI maths. The best choice depends on the task, budget, energy use and where the system must run.
An everyday example
At a service centre, one worker at the front desk handles many different tasks: greeting people, checking IDs, answering questions and deciding where each form should go. In the back room, a large team quickly compares thousands of photographed forms against known templates, each person checking one small part of many images at the same time. In a special corner, another machine is set up just for one repeated matching method and does that narrow job very fast. Together, they move forms from arrival to correct routing.
People and office machines work slowly and can change methods by discussion. Chips operate electronically at enormous speed and follow exact instruction patterns. Also, a TPU is not simply a faster GPU for every task; its advantage depends on whether the model and software can use its specific design.
Remember thisThese processors are optimised for different work: varied instructions, parallel arithmetic or specialised tensor operations. Most AI systems use them as parts of a stack, not as interchangeable upgrades.
Understand it
Understand it
Why it matters
Non-specialists often approve budgets, timelines or risk controls without realising that chip choice changes what is practical. A system that works on a CPU may be slower but easier to deploy widely. A GPU or TPU may cut response time for heavy AI tasks, but may increase cost, dependency on cloud providers or setup complexity. This affects procurement, energy use, scalability, data-location choices and whether a tool can run centrally, in the cloud or on local devices.
How CPU, GPU and TPU relate
After reading this diagram, a learner should be able to explain why different chips suit different AI and non-AI tasks, and how 'accelerator' differs from 'compute'.
01
CPU
General-purpose chip for varied instructions and system control. Good when tasks differ, branch often, or need quick coordination between many software steps.
Draw the first column with one large central block labelled 'few powerful cores' and arrows to mixed task icons: browser, spreadsheet, operating system, small AI task.
02
GPU
Chip with many smaller processing units. Good for large batches of similar calculations, such as graphics and many AI training and inference operations.
Draw the second column with a grid of many small blocks labelled 'many parallel cores' and repeated arrows to image, matrix and model icons.
03
TPU
Specialised accelerator for certain machine-learning operations, especially dense tensor maths. Strong when software and workload match the hardware design.
Draw the third column with a compact matrix engine icon and arrows to training and inference icons, plus a note 'works best for supported ML operations'.
04
How they relate
CPU, GPU and TPU are chip types with different roles. 'Accelerator' is the umbrella term for hardware that speeds up particular workloads. 'Compute' means available processing capacity across such hardware.
Place a bottom strip spanning all columns: a bracket over GPU and TPU labelled 'often called accelerators', and a separate meter icon labelled 'compute = amount of processing capacity'.
A three-column comparison shows CPU, GPU and TPU side by side. Each column follows the same read order: design focus, best at, trade-offs, and common use. A fourth small panel underneath defines accelerator as a broader category and compute as the amount of processing work, not a chip type.
On hover or tap, each column reveals one practical question: 'Do I need flexibility?', 'Do I need parallel speed?', or 'Does my model fit this specialised hardware?'
Describe this without the diagram
This comparison has three main columns and one summary strip. The CPU column says a CPU has fewer, more flexible cores and is best for mixed tasks, system control and software that changes path often. The GPU column says a GPU has many parallel cores and is best for large numbers of similar calculations, including many graphics and AI workloads. The TPU column says a TPU is a specialised accelerator for certain machine-learning maths and works best when the model and software fit its design. A final strip explains that accelerator is a broad category that often includes GPUs and TPUs, while compute means the amount of processing capacity available, not the chip itself.
With reduced motion: A static table with three columns and one footer replaces any hover effects. All labels and notes are visible at once in reading order from left to right, then the footer.
A chip is a physical part inside a computer that carries out calculations. A CPU, or central processing unit, is a general-purpose chip. It handles many kinds of instructions and controls the overall flow of work. A GPU, or graphics processing unit, has many smaller processing units that can do lots of similar calculations at the same time. A TPU, or tensor processing unit, is a more specialised accelerator designed for certain machine-learning maths.
What goes in is encoded data and instructions that may represent text, images or model values. The chip then performs operations on that input. For AI, many operations are repeated across large tables of numbers. GPUs and TPUs are often chosen for this because they can process many repeated calculations in parallel, meaning at the same time across many units.
What comes out is not only an answer such as a prediction or generated text. The choice of chip also affects how fast a system responds, how much electricity it uses, how much it costs to run, and where it can be deployed. A phone, a laptop, a data centre and a field device may use different chips for practical reasons.
These terms are related but not identical. CPU, GPU and TPU name types of chips. Accelerator is a broader term for hardware that speeds up a specific kind of work, often alongside a CPU. Compute is different again: it means the processing capacity used to train and run AI systems, not the chip itself. Cloud computing means using computing services over a network from remote data centres, instead of owning and running all the machines yourself.
How “Sorting incoming forms at a service centre” maps onto the real thing
| In the story | In the AI system |
|---|---|
| front-desk worker handling varied tasks | CPU managing general instructions and coordinating work |
| large back-room team checking many image parts at once | GPU running many similar calculations in parallel |
| special machine built for one repeated matching method | TPU or other accelerator optimised for specific machine-learning operations |
| forms being routed to the right destination | input data being processed into outputs such as classifications or predictions |
A worked example
At a regional hub, an information management officer tests a trained model that assigns already approved, non-personal project photographs to broad categories before figures are added to a monitoring report.
Starting from: A batch of cleared photographs, a trained image model and three approved computing options.
- The central processor opens and resizes the files, runs the application logic and sends prepared batches to the model. It can also run the model itself, particularly when the workload is small or infrequent.
- A graphics processor performs many of the model's matrix calculations in parallel. This can improve throughput when enough images are processed together, although the benefit depends on data transfer, batching and model design.
- A tensor processing unit executes supported tensor operations through its tuned hardware and compiler path. It may be efficient for a compatible model, but moving the model can require supported frameworks and additional setup.
- The officer compares end-to-end processing time, cost, energy use, compatibility and operational support rather than treating chip type as a speed guarantee.
Result: The preferred option depends on the workload. The central processor may be sufficient for a modest recurring report, while an accelerator may be worthwhile for larger batches. The application may still use a central processor to orchestrate work even when model calculations run elsewhere.
In a United Nations workplace
A programme officer in a country office is preparing a delegation briefing before capitals reply on a proposed image-classification pilot. The vendor's costing offers standard processors, graphics processors or access to tensor processors, each with different claims about speed and price.
The distinction helps the programme officer test the proposal rather than simply accept the largest hardware line. A central processor can coordinate the workflow and handle varied tasks, a graphics processor can accelerate parallel calculations, and a tensor processor may be efficient for supported model operations but less flexible.
Before you try thisDo not upload the briefing draft, correspondence from capitals, personal data or sensitive programme material to a service that has not been approved for that information. Selecting different hardware does not make the service secure, compliant or suitable for the data.
Where you may hear it
- A vendor demo during a procurement meeting
- An ICT governance board review of proposed cloud hosting
- A budget discussion on model training and operating costs
- A technical handover before an internal tool is deployed
Avoid confusion
Avoid confusion
Often confused with
Compute
Compute is the processing capacity used to train and run AI systems, with different needs before deployment and during everyday use.
Cloud computing
Using on-demand computing services over a network, instead of relying only on machines you manage directly.
Model hosting
The systems and services that keep an AI model available so it can receive requests, run on computing hardware, and return outputs reliably and securely.
Common misconceptions
Common misconception: A GPU is always better than a CPU for AI.
Correction: Not always. Many AI tasks involve more than the model calculation itself. Data loading, business rules, security checks and system control often still rely on CPUs. For small models or low-volume workloads, a CPU can be simpler and more cost-effective.
Common misconception: A TPU is just another name for any fast AI chip.
Correction: No. TPU usually refers to a specific specialised design for particular machine-learning operations. Other accelerators may use different architectures and software requirements. Performance depends on whether the model fits the hardware and tools.
Common misconception: More powerful chips automatically make outputs more accurate.
Correction: Faster hardware mainly changes how quickly and cheaply calculations run. Accuracy mostly depends on the model design, training data, evaluation and how well the system matches the real task.
Quick check
Which statement best explains the difference between these terms?
The key distinction is design purpose. CPUs are flexible and coordinate many kinds of tasks. GPUs speed up large numbers of similar calculations. TPUs are more specialised accelerators for particular machine-learning operations. Compute is the amount of processing capacity, not a chip name.
Go deeper
Go deeper
The simple version
Think of CPUs, GPUs and TPUs as different kinds of engines for digital work. CPUs are flexible all-rounders. GPUs are good at doing many similar calculations at once. TPUs are specialised chips for certain AI maths. The best choice depends on the task, budget, energy use and where the system must run.
The technical version
A CPU is optimised for low-latency execution of varied instruction streams, control flow and general-purpose computing. A GPU is optimised for high-throughput parallel execution across many arithmetic units, which suits matrix-heavy workloads common in graphics and machine learning. A TPU is a specialised accelerator built to execute specific tensor operations efficiently, often using hardware and compiler paths tuned for supported ML frameworks. In practice, AI systems use a stack: CPUs orchestrate data movement and application logic, while GPUs or TPUs may perform the most intensive model computations. This should be distinguished from compute, which refers to the quantity of processing capacity available over time, regardless of chip type.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
This NIST framework helps policy and leadership readers connect technical choices, including infrastructure choices, to risk management and governance responsibilities.
Intermediatelink checked 2026-07-31
Primary sourceOECD AI Principles (opens in a new tab)OECD
These OECD principles help non-technical and policy audiences place infrastructure decisions within broader responsible AI governance considerations.
Intermediatelink checked 2026-07-31
EncyclopaediaCPU cache (opens in a new tab)Wikipedia
This Wikipedia page gives a basic explanation of CPU cache, which helps beginners understand one reason CPUs handle some tasks differently from GPUs and TPUs.
Intermediatelink checked 2026-07-31
VideoVideo explanations: GPU, CPU and TPU (opens in a new tab)YouTube search
This YouTube search results page helps visual learners find multiple video explanations comparing GPUs, CPUs and TPUs, but it is a search page rather than one vetted lesson.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: GPU, CPU and TPU (opens in a new tab)Reddit search
This Reddit search results page helps curious readers see common questions and informal explanations about GPUs, CPUs and TPUs, but it is a search page with mixed-quality community discussion.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
These processors are optimised for different work: varied instructions, parallel arithmetic or specialised tensor operations. Most AI systems use them as parts of a stack, not as interchangeable upgrades.