Google launched a new computer chip called “Ironwood” last week. This is the company’s 7th generation TPU (tensor processing unit), designed to run artificial intelligence (AI) models.
This article explains what a TPU is and how it differs from other processors such as central processing units (CPUs) and graphics processing units (GPUs).
First, what is a processing unit?
A processing unit is essentially a hardware unit that is the brain of a computer. Just as the human brain processes tasks such as reading and solving math problems, processing units also perform tasks. It could be doing calculations, taking a photo, or sending a text.
What is a CPU?
A CPU was developed in the 1950s and is a general-purpose processor that can handle a variety of tasks. It’s like the conductor of an orchestra, controlling all the other components of your computer, from the GPU to the disk drive to the screen, to perform various tasks.
A CPU has at least one core, a processing unit within the CPU that can execute instructions. In the early days, CPUs only had one core, but today they can have anywhere from two to up to 16 cores. Each core of a CPU can process one task at a time, so multitasking capacity is determined by the number of cores in your hardware.
According to a report from DigitalOcean, “Two to eight cores per CPU are generally sufficient for any task an amateur needs, and the performance of these CPUs is so efficient that humans don’t even notice that tasks are being performed sequentially rather than all at once.”
What is GPU?
Unlike a CPU, a GPU is a specialized processor (a type of application-specific integrated circuit, or ASIC) that is designed to perform multiple tasks simultaneously rather than sequentially (as a CPU does). Modern GPUs are made up of thousands of cores that break down complex problems into thousands or millions of individual tasks and process them in parallel. This is a concept known as parallelism. This makes the GPU much more efficient than the CPU.
Story continues below this ad
Originally developed for graphics rendering in games and animation, GPUs are today much more flexible and are the foundation of machine learning (click here to learn more about machine learning).
According to a report on Intel’s website, “While graphics and hyper-realistic game visuals remain their primary function, GPUs have become more general-purpose parallel processors, able to perform many tasks simultaneously and handle a wide range of applications, including AI.”
This does not mean that the GPU has replaced the CPU. This is because in certain situations it is more efficient to perform tasks sequentially than to choose parallelism. Therefore, today GPUs are used as coprocessors to improve performance.
To understand this better, consider the following analogy. The CPU is like Carmy from the TV series “The Bear,” the restaurant’s head chef who has to make sure he flips hundreds of hot dog patties. Now he can do it himself, but this will be a time-consuming task. So Carmy enlists the help of Sydney, who is the sous chef and can flip the patties in parallel. A GPU is like a Sydeny with 10 hands that can flip 10 patties in 10 seconds.
Story continues below this ad
What is TPU?
TPUs are also a type of ASIC and are designed to perform a narrow range of intended tasks. First used by Google in 2015, TPUs were built specifically to accelerate machine learning workloads.
“We designed TPUs from the ground up to perform AI-based computing tasks, making them even more specialized than CPUs or GPUs. TPUs have been at the heart of Google’s most popular AI services, including search, YouTube, and DeepMind’s large-scale language models,” Chelsie Czop, a product manager who works on AI infrastructure at Google Cloud, said in an interview.
TPUs are designed to handle tensor (a general term for data structures used in machine learning) operations. Excellent at processing large amounts of data and running complex neural networks (click here Learn more about neural networks) and enable faster training of AI models. While it can take weeks to train an AI model using a GPU, the same process can be performed within hours using a TPU.
