CPU vs GPU vs TPU vs QPU: What Each Does Best

CPU: the general-purpose processor

CPU stands for central processing unit. It is the processor that runs the operating system and most everyday software.

A CPU has a small number of powerful cores. A core is one independent unit that reads instructions and executes them. Server CPUs may have dozens or a few hundred cores, and laptop CPUs have far fewer.

Each core is designed to finish one stream of instructions as quickly as it can. To do this it uses large caches, which are small, fast memories close to the core, and it predicts which way a program will branch so it can keep working without waiting. These features make a CPU good at tasks with lots of decisions, such as running a database, a web server or a compiler.

A CPU is also the most flexible chip. Almost any program can run on one. That flexibility is why every computer has a CPU even when other processors do the heavy lifting.

GPU: thousands of simple workers

GPU stands for graphics processing unit. It began as a chip for drawing images and is now the main workhorse of artificial intelligence.

A GPU has thousands of small arithmetic units. They work best when the same operation has to be applied to a very large amount of data at once. This pattern is called data parallelism. Drawing the pixels of a frame is one example. Multiplying the large grids of numbers, called matrices, that sit inside a neural network is another.

A GPU is paired with very fast memory. High-end data centre GPUs use high bandwidth memory, or HBM, which sits right next to the processor and moves data far faster than ordinary memory. Our memory guide covers this in detail.

A GPU is poor at code full of unpredictable branches, and it usually needs a CPU to run the operating system and feed it work. The software around it matters a great deal. NVIDIA's CUDA platform is a major reason its GPUs are widely used, and AMD's ROCm is the main alternative.

TPU: a chip built around matrix multiplication

TPU stands for tensor processing unit. It is the name Google gave to its own AI chips. A tensor is a grid of numbers with any number of dimensions, and a matrix is a tensor with two.

A TPU is an application-specific integrated circuit, or ASIC. That means the chip is designed for one family of jobs rather than for everything. Its core is a systolic array, a grid of multiply-and-add units where data flows from one unit to the next. This reduces how often the chip has to go back to memory, which saves time and energy.

Google first used TPUs inside its own data centres around 2015 and later offered them to cloud customers. Other companies have followed the same path with custom chips, including Amazon, Microsoft and Meta.

The trade-off is flexibility. A TPU runs the workloads it was designed for very efficiently, and it runs other workloads less well or not at all. It has mostly been available through Google Cloud, so choosing one also means accepting a particular software stack.

QPU: a different kind of machine

QPU stands for quantum processing unit. It does not work like the other three. Instead of bits that are either 0 or 1, it uses qubits, which follow the rules of quantum mechanics.

A QPU is not a faster CPU. It is a specialised accelerator that, in theory, offers large speedups on a short list of problems, such as simulating molecules and materials and factoring large numbers. For search and optimisation the gains are smaller or still unproven. For most everyday computing, such as web pages, spreadsheets or video, a QPU offers no advantage.

Today's QPUs are small and noisy. Their qubits make errors, and fixing those errors takes many physical qubits for each reliable one. A QPU also sits next to a classical computer that prepares the job and reads the result. See our guide to what a QPU is for the details.

When each one wins

Use a CPU when the work is varied, branchy or hard to split up. Operating systems, databases, web servers and most business software fall here.

Use a GPU when the same arithmetic must be repeated over huge amounts of data. Training and running neural networks, scientific simulation and graphics fall here.

Use a TPU or a similar custom AI chip when you run a stable AI workload at large scale and the cost per result matters more than flexibility.

Consider a QPU only for the narrow problems where a quantum speedup is known or plausible, and expect to experiment rather than deploy at scale. In 2026 most quantum work is research and early testing.

How they work together

Real systems combine these chips. A typical AI server has CPUs that run the operating system and move data, plus several GPUs or TPUs that do the matrix arithmetic. A quantum experiment adds a QPU, with CPUs and GPUs handling control and error decoding.

The practical question is rarely which chip is best. It is which chip is best for a given part of the job, and how fast data can move between them.