What Is a System on Chip (SoC)?

A system on chip (SoC) is an integrated circuit that puts the main parts of a computer, the CPU cores, graphics, memory controller, input and output interfaces and often an AI accelerator, onto a single chip. Where a traditional PC spreads those functions across a processor, a chipset and separate controllers, an SoC brings them together on one piece of silicon or one package.

SoCs are how most compact and low-power embedded computers are built, and they are the norm for Arm platforms. They sit inside phones, gateways, HMIs, cameras, robots and edge AI modules, and on the x86 side Intel Atom, Core Ultra and AMD Ryzen Embedded parts are highly integrated SoCs too.

This guide explains what is inside an SoC, how modern SoCs are built from dies and tiles, how an SoC compares with a socketed CPU, a system on module and a microcontroller, the advantages and trade-offs, examples of the SoCs in today’s industrial boards, and how to choose an SoC-based platform.

Rather than several chips on a board, an SoC puts the core functions on one die or package and links them with a fast on-chip interconnect. The blocks vary by chip, but most embedded SoCs include the following.

Our explainers on what a CPU is, GPU vs CPU and what an AI accelerator is describe the main blocks on their own, and our guide to what an NPU is covers the AI engine in more detail.

Most SoCs are monolithic: every block is etched on one die in one process. That gives the shortest connections and the lowest power, but the whole chip must use one process, and a large die is costly to make without defects.

Newer high-performance designs split the SoC into several dies in one package. AMD calls them chiplets and Intel calls them tiles: Intel’s Core Ultra, for example, places its compute, graphics, SoC and I/O tiles on a base die with Foveros 3D packaging, so each tile can use the process that suits it. The UCIe standard, published in 2022, aims to let chiplets from different makers work together in one package. Our guide to Intel Core Ultra shows the tiled approach in practice.

Searches for SoC vs CPU usually mean one of two things. Strictly, the CPU is one block inside an SoC: the SoC is the whole chip, and the CPU cores are the part that runs the operating system and applications. In practice, the comparison people want is between an SoC and a socketed CPU, the separate processor that sits in a socket on a desktop-style or industrial motherboard.

So is an SoC the same as a CPU? No: every SoC contains a CPU, but a CPU on its own needs other chips around it to make a computer. The line has blurred because modern x86 processors have absorbed so much, and mobile Intel and AMD parts are effectively SoCs. Socketed desktop parts still suit builds that need the most cores, PCIe slots for GPUs and capture cards, or a processor swap later, which our guide to choosing an embedded motherboard covers.

An SoC is a chip, and a chip on its own is not a product. Designing a board around a modern SoC means routing fast memory and high-speed interfaces, which takes skill and time, so most industrial products take the SoC on a module or on a ready-made board.

Our guide to the system on module explains how a module separates the compute from the carrier and why that helps with upgrades and lifecycle, our comparison of SMARC, Qseven, OSM and COM Express sets out the module standards, and our guide to what a single board computer is covers the ready-made boards. Jetson modules are a good example of the module route: the SoC reaches a product on a module that plugs into a Jetson carrier board.

A microcontroller is a small SoC in the strict sense: one chip with a CPU core, memory and peripherals. The difference is scale. A microcontroller keeps its program flash and RAM on the chip, runs firmware or a real-time OS and draws milliwatts, while an applications SoC uses gigabytes of external memory, has an MMU, boots Linux, Android or Windows and carries GPUs and NPUs.

Many SoCs carry a microcontroller-class core inside, such as the Cortex-M7 on the i.MX 8M Plus, for real-time work beside Linux. Our guide to embedded processors covers the full range of processor types.

The SoC trade is integration against flexibility. Putting everything on one chip cuts size, power and cost, and improves reliability with fewer components and fewer connections to fail. The cost is expandability: you generally cannot upgrade the CPU or add a discrete GPU to an SoC the way you can on a socketed desktop platform.

For embedded designs, where size, power and longevity matter more than upgradeability, the trade usually favours the SoC. Where a project needs several PCIe cards or the option of a faster processor later, a socketed board or a module on a carrier is the better form factor for the installation.

Most embedded SoCs use LPDDR memory soldered beside the chip, or stacked on top of it in a package-on-package layout, because the short traces let the memory run fast at low power and resist shock and vibration. Some x86 SoCs, and a few Arm ones, take SO-DIMMs instead. The memory controller on the SoC sets the type, speed and maximum capacity, so the capacity chosen when the board is ordered is usually the capacity for life.

Storage is usually soldered too. eMMC or UFS on the board holds the operating system with nothing to work loose, and an M.2 socket adds an NVMe SSD where logs, video or larger models need space. Our comparison of DDR4 and DDR5 covers the LPDDR5 side in more detail.

An SoC’s low power is its main attraction, but the heat still has to go somewhere. A 5 W Arm SoC can run fanless on a small heatsink, while a 25 W Jetson module or Core Ultra H part needs a designed heat path to the enclosure, and both slow down if the air inside gets too hot. Size the cooling for the sustained power mode the application will use, as our guide to TDP explains, and check the board’s rating against the industrial temperature range the site needs.

Lifecycle works differently too. Because everything is on one chip, an SoC reaches end of life all at once, and the boards and modules built on it follow. Industrial SoC vendors publish longevity programmes, NXP for ten or 15 years on many i.MX parts and NVIDIA for each Jetson module, and module standards soften the change by keeping the carrier when the module is replaced. Our explainer on how long NVIDIA Jetson is supported and our guide to Jetson Xavier and TX2 end of life show how a migration is planned.

These are the SoCs most often found in the industrial boards and modules we supply. Figures are the makers’ ratings, rounded, and vary by model and power mode.

Our i.MX 8M Plus and RK3588 pages list the boards on two of these SoCs, and our NVIDIA Jetson range is the SoC platform we see most for industrial edge AI, an Arm CPU paired with an NVIDIA GPU and run through the JetPack software stack. Our guides to Jetson modules and the Qualcomm Hexagon NPU cover those families in more depth.

Arm-based SoCs from NXP, TI, Qualcomm, Rockchip and others dominate low-power devices; NVIDIA Jetson modules are SoCs that add a large GPU for edge AI; and on x86, Atom, Core Ultra and Ryzen Embedded SoCs bring integration to industrial PCs. The choice between them is mostly about software and lifecycle.

Our guide to Arm architecture covers the architecture most associated with the SoC model, and our comparison of x86 and Arm sets out when each makes sense.

Single-chip designs began in the 1970s with calculators and digital watches, where one chip held the logic, memory and display drive. The modern SoC grew up in mobile phones in the 1990s and 2000s, as chip makers combined Arm cores with radio basebands, memory controllers and multimedia blocks to save board space and battery.

Industrial computing followed the same path a few years behind. The SoCs that once served phones now sit in gateways, HMIs and cameras, with industrial temperature grades and long availability added, and the edge AI modules of today are SoCs built for robots and machines rather than handsets.

SoCs are why a gateway, an HMI or an edge AI module can be compact and fanless. Embedded products value exactly what the SoC gives: a small board, low heat, fewer parts to fail and built-in engines for video, cameras and AI. They rarely need what an SoC gives up, because the I/O is fixed when the product is designed and the processor is not upgraded during its life.

Many of these SoCs reach a product on a module rather than on the main board itself, so the compute can be updated by swapping the module while the carrier and enclosure stay the same. That is the pattern behind Jetson, SMARC and COM Express designs, and it is often the best of both: SoC integration with a planned upgrade path.

These examples cover a module, three Arm boards and an x86 SoC computer. We configure the BSP or BIOS, image the operating system and application, and test each one before it ships.