CPU vs GPU vs TPU vs NPU: What's the Difference?
Every AI headline seems to introduce a new three-letter chip. CPU, GPU, TPU, NPU, and now LPU and DPU. They are not rivals so much as specialists. Here is what each one is built for, how they differ, and which one actually runs the AI you use every day.
On this page
- 01 · Why There Are So Many AI Chips Now
- 02 · CPU, GPU, TPU and NPU: Full Forms and One-Line Definitions
- 03 · The CPU: The Generalist in Charge
- 04 · The GPU: Parallel Power for Training and Inference
- 05 · The TPU: Google's Matrix-Math Specialist
- 06 · The NPU: AI on Your Own Device
- 07 · How They Work Together in Real Systems
- 08 · What About LPU, DPU and Other New Chips?
- 09 · Which Chip Should You Care About?
- 10 · Key Takeaways
- 11 · Frequently Asked Questions
01 · Why There Are So Many AI Chips Now
AI hardware explained · CPU vs GPU vs TPU vs NPU
Four letters, four very different jobs
Search for "CPU vs GPU vs TPU vs NPU" and you will find plenty of tables and very little explanation. The short answer is that all four are processors, but each is designed around a different kind of work. A CPU is a generalist that does anything, one careful step at a time. A GPU is a crowd of simple workers doing the same thing in parallel. A TPU is a machine built almost entirely for the matrix math at the heart of AI. An NPU is a small, power-sipping AI engine tucked inside your laptop or phone.
The reason this matters now is scale. Modern AI models are trained on enormous datasets and then asked to answer millions of questions a day. Doing that math on general-purpose chips is slow and expensive, so the industry has built specialized accelerators for each part of the job. Understanding which chip does what helps you pick the right laptop, the right cloud instance, and the right career skills.
02 · CPU, GPU, TPU and NPU: Full Forms and One-Line Definitions
Answer: CPU stands for Central Processing Unit, GPU for Graphics Processing Unit, TPU for Tensor Processing Unit and NPU for Neural Processing Unit. The CPU is a general-purpose processor, the GPU a massively parallel processor, the TPU Google's matrix-math accelerator, and the NPU a low-power AI engine for devices.
One quick warning before going further: if you search for "TPU" on its own, half the results are about thermoplastic polyurethane, the flexible plastic used in phone cases. In computing, TPU always means Google's Tensor Processing Unit. With that settled, here are the four chips in one table.
| Chip | Full form | One-line definition | Made by (examples) |
|---|---|---|---|
| CPU | Central Processing Unit | General-purpose processor that runs any program step by step | Intel, AMD, Arm-based designs, Apple |
| GPU | Graphics Processing Unit | Thousands of simple cores doing the same math in parallel | NVIDIA, AMD, Intel |
| TPU | Tensor Processing Unit | Custom chip built around matrix multiply units for AI | |
| NPU | Neural Processing Unit | Low-power AI accelerator built into laptop and phone chips | Qualcomm, Intel, AMD, Apple (Neural Engine) |
The figure below puts the same comparison side by side across five questions: full form, what each chip is built for, what is inside it, what it is best at, and where you will find it.
Read across the "Inside" row and the whole story makes sense. Few powerful cores make the CPU flexible. Thousands of simple cores make the GPU parallel. Dedicated matrix units make the TPU efficient at AI math. Small, efficient AI engines let the NPU run all day on battery.
03 · The CPU: The Generalist in Charge
Answer: A CPU has a small number of powerful cores optimized for running varied, step-by-step logic quickly. It runs the operating system, your apps and the code that feeds work to accelerators. It can run small AI models, but it is slow and inefficient for large ones.
Think of the CPU as a skilled manager. It is excellent at decisions, branches and juggling many different tasks: opening files, handling network traffic, running a database query, deciding what happens next. Modern server CPUs have dozens or even over a hundred cores, but each core is built to be smart and flexible rather than to repeat one calculation millions of times.
That flexibility is why every computer has a CPU, including every AI server. Even in a data center packed with GPUs or TPUs, host CPUs load data, preprocess it, schedule jobs and handle everything that is not raw matrix math. For AI, CPUs are perfectly fine for small models, classic machine learning and light inference, which is why cloud providers including OCI still offer plenty of CPU instances for AI-adjacent work.
04 · The GPU: Parallel Power for Training and Inference
Answer: A GPU contains thousands of simpler cores that run the same operation on many pieces of data at once. That design, built for graphics, turned out to be ideal for neural networks, which is why GPUs became the default chip for AI training and inference.
A neural network is mostly multiplying and adding large grids of numbers. A GPU can split that work across thousands of cores and finish it far faster than a CPU. Data-center GPUs add large amounts of high-bandwidth memory (HBM) so whole models fit close to the compute, and fast links between GPUs so a single model can be spread across many of them.
Why GPUs dominate today
Hardware is only half of it. NVIDIA's CUDA software platform, plus libraries tuned for deep learning, made GPUs easy to program, and almost every AI framework supports them first. AMD offers an alternative software stack called ROCm. The result is a huge ecosystem: if a new model is released, it almost certainly runs on GPUs on day one. You can rent GPUs in every major cloud, including NVIDIA and AMD Instinct instances on OCI.
05 · The TPU: Google's Matrix-Math Specialist
Answer: A TPU, or Tensor Processing Unit, is a custom chip Google designed for AI. Instead of thousands of general cores, it is built around matrix multiply units called systolic arrays, so it spends almost all of its silicon and power on the math neural networks need.
Google started building TPUs over a decade ago for its own services and later offered them to customers through Google Cloud. A systolic array works like a production line: numbers flow through a grid of small multiply-and-add cells, each passing its result to the next, so data is reused instead of being fetched again and again from memory. That makes TPUs very efficient at large, regular AI workloads.
The latest TPUs
In April 2026 Google split its eighth TPU generation into two chips. Google says TPU 8t, for training, delivers nearly 3x the compute per pod of the previous generation, with superpods of up to 9,600 chips. TPU 8i, for inference, has 288 GB of high-bandwidth memory per chip and pods of 1,152 chips. On 27 October 2026, Oracle announced that Google TPUs will be available on OCI from calendar 2027, the first hyperscale cloud outside Google to offer them. ExaGuru's "Google TPUs on OCI" series covers that announcement in depth.
TPUs are programmed through frameworks rather than directly: JAX natively, PyTorch through Google's TorchTPU runtime, and the XLA compiler underneath. Code written for CUDA does not run on TPUs, which is the main trade-off to weigh.
06 · The NPU: AI on Your Own Device
Answer: An NPU, or Neural Processing Unit, is a small AI accelerator built into the main chip of laptops and phones. It runs AI features such as background blur, live captions and on-device assistants using far less power than the CPU or GPU, which saves battery and keeps data on the device.
NPUs exist because running AI on a CPU or GPU drains a battery quickly. An NPU is designed to do modest amounts of AI math very efficiently, all day. That is what powers features like noise suppression on video calls, photo enhancement and the new wave of "AI PC" tools.
What is NPU TOPS?
NPU performance is usually quoted in TOPS: trillions of operations per second, typically measured at low precision (INT8). Microsoft's Copilot+ PC program requires an NPU of at least 40 TOPS. Qualcomm's Snapdragon X Elite NPU is rated at 45 TOPS and the newer Snapdragon X2 Elite family at 80 TOPS; Intel's Core Ultra 200V and AMD's Ryzen AI 300 NPUs are rated at up to 48 and 50 TOPS, and Apple's M4 Neural Engine at 38 TOPS. Vendors measure TOPS differently, so treat them as a rough guide, not a race.
NPUs are excellent for small, always-on models. For larger local models, such as a chatbot running on your laptop, a GPU with more memory is usually faster. Post 2 in this series, "NPU vs GPU", covers that choice in detail.
07 · How They Work Together in Real Systems
Answer: Real AI systems use several chip types at once. The CPU coordinates, GPUs or TPUs train and serve large models in the cloud, NPUs run lighter AI on your device, and DPUs in data-center servers move data so the other chips stay busy.
It is tempting to ask which chip is best, but in practice they are a team. The figure below shows a typical setup: on the left, a laptop whose system-on-chip combines a CPU, an NPU, an integrated GPU and shared memory; on the right, a cloud region where host CPUs hand training jobs to GPU clusters or TPU pods, and the trained model is deployed as an inference API.
Follow the arrows. Light AI stays on the laptop's NPU. Bigger requests travel to the cloud, where a GPU cluster or TPU pods do the heavy lifting. The red dashed arrow is the CPU handing jobs to accelerators, and the grey arrow shows that either accelerator can often do the same job.
One AI feature, four chips
The same teamwork shows up across a model's life. The next figure follows one model from raw data to your device, and then uses a meeting app with live captions as a concrete example.
Notice that no single chip appears in every step. Data preparation is CPU work, training needs a GPU or TPU, serving can use GPUs, TPUs or specialist chips like LPUs, and the final on-device step belongs to the NPU.
08 · What About LPU, DPU and Other New Chips?
Answer: An LPU is Groq's Language Processing Unit, built for very fast AI inference by keeping model weights in large on-chip memory. A DPU, or Data Processing Unit, offloads networking, storage and security work from server CPUs. Both are specialists rather than replacements for CPUs, GPUs or TPUs.
| Chip | Full form | What it does | Where you meet it |
|---|---|---|---|
| LPU | Language Processing Unit (Groq) | Fast, predictable inference for language models using on-chip SRAM | Inference services; in Dec 2025 NVIDIA reportedly struck a large licensing and hiring deal with Groq |
| DPU | Data Processing Unit | Handles network, storage and security tasks so CPUs and GPUs stay free | Data-center servers, for example NVIDIA BlueField |
| ASIC | Application-Specific Integrated Circuit | Any chip custom-built for one job; TPUs and LPUs are AI ASICs | Cloud providers' custom AI chips |
The pattern is clear: as AI grows, the industry builds more specialists. The four core chips, CPU, GPU, TPU and NPU, still cover almost everything most developers and IT teams will touch.
09 · Which Chip Should You Care About?
Your answer depends on what you do. Use this as a quick guide.
- Buying a laptop: look for a capable NPU (40+ TOPS for Copilot+ features) and enough RAM; add a discrete GPU if you want to run larger local models.
- Learning AI or data science: start on CPUs and free cloud GPUs or TPUs (for example notebook services), then learn one framework well, such as PyTorch or JAX.
- Building AI products: design for GPUs today, keep code at the framework level so TPUs and other accelerators stay an option, and push light features to the NPU.
- Running cloud infrastructure: plan for a mix: CPUs for services and data, GPUs for most AI, TPUs (on Google Cloud, and on OCI from 2027) for large-scale workloads.
The best AI chip is the one that matches the job. Most real systems use all of them.
10 · Key Takeaways
- CPU is the generalist. It runs everything and coordinates the accelerators, but is slow for large AI models.
- GPU is the default AI workhorse. Thousands of parallel cores and a huge software ecosystem make it the safest choice for training and inference.
- TPU is Google's AI specialist. Built around matrix units, available in Google Cloud and on OCI from 2027, programmed through JAX and PyTorch.
- NPU brings AI to your device. Low power, measured in TOPS, ideal for always-on features and small models.
- They work as a team. Real AI products combine several chips; LPUs and DPUs are newer specialists for inference and data movement.
11 · Frequently Asked Questions
What is the difference between CPU, GPU, TPU and NPU?
A CPU is a general-purpose processor that runs any program step by step. A GPU has thousands of cores for parallel work and is the main chip for AI training. A TPU is Google's custom chip built around matrix units for AI. An NPU is a low-power AI engine inside laptops and phones.
What are the full forms of CPU, GPU, TPU and NPU?
CPU is Central Processing Unit, GPU is Graphics Processing Unit, TPU is Tensor Processing Unit and NPU is Neural Processing Unit.
Is a TPU better than a GPU?
Neither is better for everything. TPUs can be very efficient for large training and inference jobs written in JAX or PyTorch, while GPUs offer the broadest software support and availability. The right choice depends on your code, scale and where you run it.
Do I need an NPU in my laptop?
An NPU of 40 TOPS or more is needed for Microsoft Copilot+ PC features and helps run AI features efficiently on battery. For running larger AI models locally, GPU memory matters more than NPU TOPS.
What is an LPU?
An LPU is Groq's Language Processing Unit, a chip designed for very fast AI inference that keeps model weights in large on-chip memory instead of external memory.
Learn this hands-on
Practice Exadata, OCI and multi-cloud architectures in live labs with a practicing consultant.
More in AI hardware
NPU vs GPU for AI: Which Is Better for Laptops and Local LLMs?
What an NPU in your laptop actually does, what TOPS means, and when a GPU is still the better choice for running AI locally.
AI acceleratorsTPU vs GPU: Which Is Better for AI Training, Inference and Cost?
How Google TPUs and GPUs compare on performance, cost, software and availability, and how to choose for your AI workload.
AI hardware explainedCPU vs GPU vs TPU vs NPU: What's the Difference?
The four processors behind modern AI, what each is built for, and where LPUs and DPUs fit, explained without the jargon.