AI Basics: CPU, GPU and TPU

AI Basics: CPU, GPU and TPU Latest News

How TPU is Different from CPU and GPU

About Processing Units

• Processing units are the core hardware components that act as the brain of a computer. They perform tasks such as calculations, processing images, and transmitting data, similar to how the human brain executes complex tasks.

‹ ›

What is a CPU (Central Processing Unit)?

• The CPU is a general-purpose processor that was developed in the 1950s and can handle a wide variety of tasks.
• It functions like a conductor in an orchestra, coordinating the operations of all other computer parts like GPUs, disk drives, and memory units.
• A CPU contains cores — individual units that execute instructions. Early CPUs had only one core, but modern CPUs may contain 2 to 16 cores.
• Each core can handle one task at a time, so a CPU’s multitasking capacity depends on the number of cores.
• For everyday users, 2 to 8 cores are usually sufficient, and CPUs are so efficient that users rarely notice that tasks are completed sequentially, not simultaneously.

What is a GPU (Graphics Processing Unit)?

• A GPU is a specialised processor designed to perform many tasks simultaneously, using a technique called parallel processing.
• Unlike CPUs, which process tasks sequentially, GPUs break down complex tasks into thousands or millions of smaller problems, solving them in parallel.
• Modern GPUs contain thousands of cores, making them far more suitable for intensive computational tasks.
• Initially developed for rendering graphics in gaming and animation, GPUs are now widely used in machine learning and artificial intelligence.
• GPUs have evolved into general-purpose parallel processors, making them a key tool in running AI models and handling large data operations.
• However, GPUs have not replaced CPUs, because certain operations are better handled sequentially, which is the strength of CPUs.

What is a TPU (Tensor Processing Unit)?

• A TPU is also a type of ASIC (Application-Specific Integrated Circuit), meaning it is built for a specific function — in this case, AI tasks.
• First introduced by Google in 2015, TPUs are specially designed hardware units built from the ground up to handle machine learning operations.
• TPUs focus on processing tensors — the multidimensional data arrays used in AI model computations.
• They are optimised to run neural networks efficiently, enabling faster training and execution of AI models than GPUs or CPUs.
• For example, training an AI model that may take weeks on a GPU can often be completed in hours using a TPU.
• TPUs are used at the core of Google’s major AI services, such as Search, YouTube, and DeepMind’s large language models, illustrating their real-world application in high-scale AI infrastructure.

Source: IE