As the world focuses on harnessing the newest wave of AI technologies, one piece of high-tech hardware has grow to be a surprisingly popular commodity: the graphics processor, or GPU.

A top-of-the-line GPU will be sold Tens of 1000’s of dollarsand the leading manufacturer NVIDIA has seen its market valuation rise to over $2 trillion as demand for its products increases.

GPUs aren’t just high-end AI products either. There are also less powerful GPUs in phones, laptops and game consoles.

You’re probably wondering: what exactly is a GPU? And what makes them so special?

What is a GPU?

GPUs were originally designed primarily for quickly generating and displaying complex 3D scenes and objects, akin to those present in video games computer-aided design Software. Modern GPUs also tackle tasks akin to Decompress Video streams.

The “brains” of most computers is a chip called a central processing unit (CPU). CPUs will be used to generate graphical scenes and decompress videos, but they’re typically far slower and fewer efficient at these tasks in comparison with GPUs. CPUs are higher fitted to general computing tasks akin to word processing and browsing web pages.

How are GPUs different from CPUs?

A typical modern CPU consists of 8 to 16 inchescores“, each of which might process complex tasks one after the opposite.

GPUs, however, have 1000’s of relatively small cores which are designed to work all at the identical time (“parallel”) to attain fast overall processing. This makes them well fitted to tasks that require a big number of straightforward operations that will be performed concurrently reasonably than sequentially.

Traditional GPUs are available in two predominant flavors.

First, there are standalone chips which are often included in add-on cards for giant desktop computers. Second, they’re GPUs combined with a CPU in the identical chip package, commonly present in laptops and gaming consoles just like the PlayStation 5. In each cases, the CPU controls what the GPU does.

Why are GPUs so useful for AI?

It seems that GPUs will be used for greater than just generating graphical scenes.

Many of the machine learning techniques behind artificial intelligence (AI), akin to: deep neural networksrely heavily on various types of “matrix multiplication”.

This is a mathematical operation that involves multiplying and summing very large sets of numbers. These operations lend themselves well to parallel processing and might subsequently be performed in a short time by GPUs.

What’s next for GPUs?

The computing power of GPUs is always increasing as a result of the increasing variety of cores and their operating speeds. These improvements are primarily as a result of improvements in chip manufacturing by firms akin to: B. attributed TSMC in Taiwan.

The size of individual transistors – the essential components of each computer chip – is decreasing, allowing more transistors to be packed into the identical physical area.

However, that is not the entire story. While traditional GPUs are useful for AI-related computing tasks, they should not optimal.

Just as GPUs were originally designed to hurry up computers by providing specialized graphics processing, there are accelerators designed to hurry up machine learning tasks. These accelerators are also known as “data center GPUs.”

Some of the preferred accelerators, made by firms like AMD and NVIDIA, were originally traditional GPUs. Over time, their designs have evolved to higher handle various machine learning tasks, akin to supporting more efficient “Brain floating” Number format.

NVIDIA’s latest GPUs have special features to speed up the “Transformer” software utilized in many modern AI applications.

Other accelerators like Google’s Tensor processing units and Tenstorrents Tensix coreswere designed from the bottom as much as speed up deep neural networks.

Data center GPUs and other AI accelerators typically have significantly more memory than traditional GPU add-in cards, which is critical for training large AI models. The larger the AI ​​model, the more powerful and accurate it’s.

To further speed up training and handle larger AI models like ChatGPT, many data center GPUs will be combined right into a supercomputer. This requires more complex software to properly utilize the available number processing power. Another approach is to create a single very large accelerator, akin to “Wafer-scale processor” produced by Cerebras.

Are specialty chips the long run?

The CPUs don’t stand still either. Current CPUs from AMD and Intel have built-in low-level instructions that speed up the number processing required by deep neural networks. This additional functionality is especially helpful for “inference” tasks – i.e. the usage of AI models which have already been developed elsewhere.

In order to have the ability to coach the AI ​​models in any respect, large GPU-like accelerators are still needed.

It is feasible to create increasingly specialized accelerators for specific machine learning algorithms. For example, recently an organization called Groq launched a “Speech processing unit” (LPU), specifically designed to run large language models modeled on ChatGPT.

However, the event of those specialized processors requires significant technical resources. History shows that the usage and recognition of a given machine learning algorithm tends to peak after which decline – so expensive specialized hardware can quickly grow to be obsolete.

However, this should not be an issue for the common consumer. The GPUs and other chips within the products you employ are probably quietly getting faster.

This article was originally published at