iGPU vs dGPU: What Is the Difference?

An iGPU, or integrated GPU, is a graphics processor built onto the same chip as the CPU. It has no memory of its own and borrows part of the system RAM. A dGPU, or discrete GPU, is a separate graphics processor on its own card or module, with its own dedicated memory, power delivery and cooling. People also call them integrated and dedicated graphics.

Both do the same kinds of work: drive displays, decode and encode video, and run parallel maths such as AI inference. The difference is scale. A discrete GPU has many more cores, much faster memory and a far larger power budget, so it is much faster on heavy work, at the cost of power, heat, space and price. Our explainer on graphics cards covers what sits on a discrete card.

A discrete graphics card is a dGPU on its own plug-in board: a graphics processor with its own memory (VRAM), power circuitry and cooler, fitted in a PCIe slot on the motherboard. “Discrete” means separate from the processor, and a dedicated graphics card is the same thing under another name. Smaller cards draw all their power from the slot; larger ones also need power cables from the supply.

Not every discrete GPU is a card. In laptops, compact embedded boards and rugged computers the same class of chip is often soldered to the board or fitted as a plug-in module, and it still counts as discrete graphics because it has memory of its own. For industrial builds, low-profile and single-slot workstation cards are common because they fit compact chassis and are supported for longer than consumer gaming cards. Our GPU cards range covers PCIe graphics cards and MXM modules for industrial systems.

The table sets the two side by side for an industrial or embedded computer.

Integrated graphics have moved a long way. Intel’s Iris Xe and the Arc graphics in Core Ultra processors, and AMD’s Radeon graphics in Ryzen processors, can drive several 4K displays, decode many video streams in hardware and run compact vision models. Many current processors also add an NPU beside the iGPU, so inference can be split between the two.

Software matters as much as silicon here. Intel’s OpenVINO toolkit compiles models to run on the CPU, iGPU and NPU together. The COM Express reverse vending case study shows an Iris Xe iGPU running deep learning in a compact system with no discrete card.

A discrete GPU earns its power and heat when the workload needs more than shared memory and a processor’s power budget can give. These are the usual signs.

Desktop graphics cards are not built for sealed cabinets, vibration or wide temperature ranges, so industrial systems take discrete GPUs in three main forms. Full-height or low-profile PCIe cards fit in our industrial GPU computers, with ducted airflow sized for the card. MXM GPU modules put a discrete GPU on a compact board that can be conduction-cooled in a rugged or fanless chassis. Some embedded boards also solder the GPU directly onto the board for vibration-heavy sites.

NVIDIA Jetson sits between the two camps. Its GPU is integrated and shares memory with the CPU, like an iGPU, but it is designed for AI first and delivers far more inference performance per watt than a general-purpose iGPU. For many edge AI builds, a Jetson module is the alternative to adding a discrete card.

Start from the workload rather than the hardware. Count the cameras or video streams, note the resolution and frame rate each one needs, and list the models you plan to run. Then check the power budget and the cooling the site allows: a sealed fanless box on a 24 V supply has very different limits from a ventilated rack. If an iGPU, with or without the processor’s NPU, can run the workload with headroom, it gives the smallest, quietest and longest-lived build. If not, a Jetson module or a discrete GPU is the next step. Our edge AI computers cover all three routes.

We specify edge AI computers with integrated graphics, discrete GPUs or Jetson modules to match the workload, and fit and validate the GPU and drivers as standard before the system ships.