What is accelerated computing?

Accelerated computing depends on a wide range of hardware and software (also known as accelerators), including graphics processing units (GPUs), application-specific integrated circuits (ASICs) and field programmable gate arrays (FPGAs). 

Accelerated computing solutions are in high demand across many industries because they can perform calculations faster and more efficiently than traditional central processing units (CPUs). Unlike CPUs, accelerators rely on parallel computing, a method of computational problem solving in which tasks are broken down into smaller problems and solved simultaneously, rather than serially. 

Because of its data processing speeds, accelerated computing has become critical to the advancement of many cutting-edge technologies and applications, including artificial intelligence (AI), generative AI, machine learning (ML) and high-performance computing (HPC). Today, it is a key component of the strategies of many of the world’s most successful tech companies, including Google, Amazon Web Services (AWS) and Microsoft.

Central processing, units or CPUs, consist of various electronic circuitries that run a computer’s operating system (OS) and apps. For many years, the CPU served as the brain of a computer, transforming data input into information output. However, as applications became more advanced, they needed to process data faster and more efficiently than CPUs could manage. Enter accelerators and accelerated computing technologies with their parallel processing capabilities, low latency and high throughput. Since the 1980s, when they first gained prominence, many of the biggest technological advancements in computer science have depended on accelerators.

From the most exciting new video games and immersive virtual reality (VR) experiences to ChatGPT, the training of AI models and big data analysis, accelerators are an essential part of our rapidly evolving, hyperconnected world. Many modern enterprises rely on accelerators to power their most valuable applications and infrastructure architectures, including cloud computing, data centers, edge computing and large language models (LLMs). For example, business leaders and developers looking to explore generative AI are investing in accelerators to help optimize their data centers and process more information faster¹.

Accelerators are used across a wide range of business applications to speed data processing—especially as 5G coverage expands—increasing Internet of Things (IoT) and edge computing opportunities. IoT applications depend on accelerators to process data from smart devices like refrigerators, traffic flow sensors and more. Edge computing can deliver deeper insights, faster response times and improved customer experiences, but only with the processing speeds that accelerators offer.  

When it comes to AI, many of its most advanced applications—such as natural language processing (NLP), computer vision and speech recognition—rely on the power of accelerated computing to function. For example, neural networks that underpin many cutting-edge AI applications need AI accelerators to classify and cluster data at a high velocity.

Lastly, as more businesses seek ways to digitally transform and accelerate innovation, accelerated computing solutions offer a relatively low total cost of ownership. Accelerators’ ability to process large amounts of data swiftly and accurately means that they can be used across many different applications with the potential to create business value, including AI chatbots, financial data analytics, cloud computing and more.

Accelerated computing uses a combination of hardware, software and networking technologies to help modern enterprises power their most advanced applications. Hardware components that are critical to accelerators include GPUs, ASICs and FPGAs. Software and application programming interfaces (APIs) are just as important, with CUDA and OpenCL playing important roles. 

Finally, networking solutions, such as PCI express (PCIe) and NV link, help processing units communicate with the memory and storage devices where data is kept. Here’s a closer look at how hardware accelerators, software accelerators and network solutions work together to make accelerated computing possible.

Modern hardware accelerators can process data significantly faster than traditional CPUs due to their parallel processing capabilities. Without them, many of accelerated computing’s most important applications wouldn’t be possible.

GPUs

GPUs, or graphics processing units, are hardware accelerators designed to speed computer graphics and image processing on various devices, including video cards, system boards, mobile phones and personal computers (PCs). GPU accelerators significantly reduce the amount of time a computer needs to run multiple programs. GPU-accelerated computing is used in a wide range of accelerated computing applications including AI and blockchain.

ASICs

ASICs, or application-specific integrated circuits, are hardware accelerators built with a specific purpose or function in mind—like deep learning, in the case of the WSE-3 ASICs accelerator, considered to be one of the fastest AI accelerators in the world². Unlike some other hardware accelerators, ASICs cannot be reprogrammed. However, because they are constructed with a singular purpose, they typically outperform accelerators that have been built for more general-purpose compute tasks. Another example of an ASICs accelerator is Google’s Tensor Processing Unit (TPU), which was developed for neural network ML on Google's own TensorFlow software.

FPGAs

FPGAs, or field programmable gate arrays, are highly customizable AI accelerators that depend on specialized knowledge to be reprogrammed for a specific purpose. Unlike other hardware accelerators, FPGAs have a unique design that suits a specific function, often having to do with real-time data processing. FPGAs are reprogrammable on a hardware level, enabling a much higher level of customization. They are often used in aerospace, IoT applications and wireless networking solutions. 

APIs and software play critical roles in making accelerators function, interfacing between the hardware and the networks that are needed to run accelerated computing applications. 

APIs

APIs, or application programming interfaces, are sets of rules that allow applications to communicate and exchange data. APIs are critical to accelerated computing, helping integrate data, services and functionality between applications. They simplify and accelerate application and software development by allowing developers to integrate data, services and capabilities from other applications and enabling them to be shared across an entire organization. APIs help optimize data flow between hardware and software accelerators and give developers access to software libraries that are critical to app and software development.

CUDA

Compute Unified Device Architecture (CUDA), built by NVIDIA in 2007, is software that gives developers direct access to NVIDIA GPUs’ parallel computation abilities. CUDA empowers coders to use GPU technology for a much wider range of functions than was previously possible. Since then, building on what CUDA made possible, GPU hardware accelerators have gained even more capabilities—perhaps most significantly ray tracing, the generation of computer images by tracing the direction of light from a camera, and tensor cores that enable DL.

OpenCL

OpenCL is an open source platform designed for parallel computing and supports many kinds of hardware accelerators, including GPUs and FPGAs. Its high compatibility makes it an ideal tool for developers who need to use different types of components in their accelerated computing workloads. Examples of OpenCl use cases include gaming, 3D modeling and multimedia production.

Networking technologies are critical to accelerated computing, enabling fast, effective communication between the many different processing units and memory and storage devices where data is being stored. Here are some of the different types of networking that accelerated computing relies on.

Ethernet

Ethernet is a kind of technology that is widely used to provide fast, flexible transfer of data between servers in a data center (or just between computers that are in the same physical space). While it’s widespread and affordable, it isn’t as fast as some of the other kinds of networking, like NVLink or InfiniBand.

PCI Express (PCIe)

PCIe is a high-speed computer expansion bus that connects two devices with an external memory source. Accelerators use PCIe to connect GPUs or other kinds of hardware accelerators to a central computing system.

NVLink

NVLink is NVIDIA’s proprietary interconnect technology and can deliver much higher bandwidth than PCIe. It was built to enable highly efficient data sharing between GPUs and other devices.

InfiniBand

InfiniBand is a communications specification that defines switched fabric architecture on interconnected servers, storage or other devices in a data center. Built by the InfiniBand Trade Association, the technology is distinguished by its high performance and low latency, making it ideal for high-performance workloads.

Computer Express Link (CXL)

CXL is an open interconnect standard that helps achieve low latency and increase bandwidth between CPUs and accelerators by combining several interfaces into a single PCIe connection. 

With the spread of AI technology and the expansion of 5G networks enabling lightning-fast data transfer speeds, the number of accelerated computing use cases is growing every day. Here are some of the most common.