CPU vs. GPU

The CPU (Central Processing Unit) was the core of any computer throughout decades of its history. We referred to it as the brain, the quick, intelligent and universal beaver that powered our operating systems and applications. The GPU (Graphics Processing Unit) was considered an expert in the corner, a one-trick pony that only plays with video games, making them more beautiful. Then there was Artificial Intelligence and all was changed. The whole AI hype, with complex chatbots and scientific studies on the subject, is occurring nearly solely due to our discovery that the GPU in question had the one trick. It happens that AI training requires a form of mathematics to which CPUs are horrifyingly inefficient, and which GPUs are made to do.

The CPU - The Master Chef A CPU is a serial processor. Imagine it to be a head chef in a fine restaurant. It possesses some, very strong cores (the hands of the chef) which may do any complicated, sequential action very fast. It has the ability to move out of running your web browser to computing a formula in the spreadsheet to access your files. It is a brilliant generalist. But shall this great cook cut 10,000 carrots a quarter? It can't. It will cut them, one after another, very fast, still it is a huge bottleneck.

The Massive Kitchen Crew The GPU. A GPU is a parallel processor. Imagine it to be a giant kitchen with 1,000 line cooks. It was meant to render graphics i.e. calculate the color and position of millions of pixels at once per frame. Individually each of the cooks (or cores) is far less complex than the master cook but there are thousands of them. They are not able to run your operating system, but are able to chop one carrot all together, at precisely the same time.

The AI "Aha!" Moment

So, what is "training an AI"? Deep learning essentially is a colossal sequence of mathematical computations, namely, matrix multiplication. It is merely billions and billions of mere basic, simple multiply and add computations.

Our master cook, the CPU, would need months or years to accomplish this, calculation by calculation. The G.P.U. consisting of 1,000s of cooks is capable of making thousands of them simultaneously. This was named GPGPU (General-purpose computing on GPUs) and reduced AI training times by a factor of years, days, or hours.

It is still the CPU that is the brain of the show. However, the key component is the GPU, which does the strenuous work of the AI revolution. It is not about the future of a single one of them being better but about the strong alliance between the serial master and the parallel army.