CPU vs GPU: Why Computers Use Two Kinds of Processor

A CPU (central processing unit) and a GPU (graphics processing unit) are both chips that do calculations, but they are built for different jobs. A CPU has a small number of strong cores and is designed to finish each task quickly. A GPU has a very large number of simpler cores and is designed to push a lot of similar work through at the same time. Most computers, phones and servers use both, and each does the work it suits best.

What a CPU does

The CPU is the main processor of a computer. According to Wikipedia's article on the subject, it runs the instructions of a program and performs the arithmetic, logic, control and input/output operations those instructions ask for. Inside it, an arithmetic logic unit does the sums and bit-level logic, registers hold the numbers being worked on, and a control unit directs the steps.

Those steps repeat in a loop called the instruction cycle. The CPU fetches the next instruction from memory, decodes it to see what operation it asks for, and executes it. The result goes to a register or back to memory, and the cycle begins again. Because a program often has to make decisions along the way (if this, then that), CPUs include complex control logic for branching and scheduling.

CPUs also carry caches, small and very fast memories close to each core that keep copies of data used often. Modern chips usually have several levels: a small, fast L1 cache per core, a larger L2, and a larger L3 that cores typically share. A multi-core CPU puts several cores on one chip so it can run several threads or programs at once. GeeksforGeeks puts typical CPU core counts at roughly 4 to 64.

What a GPU does

A GPU began as a chip for graphics. Wikipedia describes it as a specialised circuit designed to accelerate computer graphics and image processing. Drawing a scene on a screen means working out the colour of millions of pixels, and the calculation for one pixel is largely independent of the next. That makes the job easy to split across many small calculation units, and modern GPUs have, in Wikipedia's words, hundreds or thousands of them.

Intel's comparison page describes the GPU's style as parallel: it divides work across many cores at once. Each core is simpler than a CPU core, and the control logic and caches are smaller. Instead of relying on big caches, GPUs rely on high memory bandwidth, meaning they can move large amounts of data in and out quickly. Problems that split neatly into many identical pieces are sometimes called "embarrassingly parallel", and they are where GPUs shine.

Latency versus throughput

Two words help here. Latency is how long one task takes from start to finish. Throughput is how much total work gets done in a given time. A CPU is tuned for low latency: when you click a button, you want that one action handled at once. A GPU is tuned for high throughput: when you draw a frame, you want millions of small calculations done together, and it matters little how long any single one takes.

• CPU: few strong cores, complex control logic, large caches, good at varied and step-by-step tasks
• GPU: many simple cores, simpler control logic, high memory bandwidth, good at repeating the same operation on lots of data

Typical jobs

According to Intel, CPUs handle running the operating system and programs, web browsing, spreadsheets, databases and game physics calculations. GPUs handle rendering graphics in games, video editing and AI workloads. GeeksforGeeks adds matrix and vector maths and image processing to the GPU list.

The AI link is worth a closer look. Training and running neural networks involves enormous amounts of multiplication and addition on grids of numbers, called matrices. Wikipedia notes that GPUs are increasingly used for artificial intelligence because they accelerate this kind of linear algebra, and that Nvidia added dedicated Tensor cores for deep learning in its Turing design in 2018.

From graphics chip to general-purpose helper

Early graphics chips were fixed in what they could do. Over time they became programmable. Wikipedia's timeline mentions Texas Instruments' TMS34010 in 1986 as described as the first fully programmable graphics processor, and Nvidia's CUDA platform in 2007 as the earliest widely adopted model for general-purpose computing on a GPU. Using a GPU for tasks other than graphics is called GPGPU. Wikipedia gives the protein-folding project Folding@home as an example, saying it can run up to forty times faster on GPUs than on CPUs in some circumstances.

How they work together

In a typical program the CPU is in charge. It runs the main logic, prepares the data and hands the heavy, repetitive part to the GPU. The GPU does its calculation and returns the result. Intel's page concludes that the question is no longer CPU versus GPU, because most workloads benefit from both.

Chip makers also combine the two. Some CPUs include a GPU on the same chip for built-in graphics, which suits laptops, tablets and phones where size and energy use matter, according to Intel. Others add a neural processing unit (NPU), a block designed to help with AI inference, the stage where an already trained model makes predictions. A separate, discrete graphics card adds more parallel power when a task needs it.

Reading specifications with care

Because the designs differ, comparing a CPU's core count with a GPU's is not meaningful. A GPU core is far simpler, so thousands of them are not the same as thousands of CPU cores. GPU speed is often quoted in teraflops (trillions of floating-point operations per second), which Wikipedia describes as an estimate. A figure like this says little about how a chip will perform on a specific program. General information only; check independent test results for the workload you care about.

Sources

• Central processing unit, Wikipedia
• Graphics processing unit, Wikipedia
• CPU vs. GPU, Intel
• Difference between CPU and GPU, GeeksforGeeks