US What makes a GPU embedded
An embedded GPU is less a different kind of chip than a different way of building and supporting one. The graphics processor is often the same silicon as a desktop or laptop part, but the embedded version is packaged for systems that are designed once and built for years: a fixed board layout, a defined power and heat limit, long availability and drivers maintained for long-lived operating systems.
That matters to anyone building machines, vehicles, medical equipment or defense systems. A desktop graphics card may change or disappear within a product cycle, forcing a redesign and fresh approvals. An embedded GPU is planned to stay available for much longer, so a qualified design can keep shipping. Searches for embedded graphics often mean something narrower, the graphics built into an embedded processor, which is one of the four types below.
The four main types
Embedded GPUs reach industrial systems in four main forms:
-
//
MXM GPU modules. A discrete GPU and its memory on a standard module that lies flat on a carrier and cools through the chassis. It is the most common discrete GPU in rugged systems; see what an MXM GPU is.
-
//
Embedded PCIe cards. Low-profile or single-slot graphics cards built on embedded GPUs, for industrial and box PCs with an expansion slot. Some GPUs, such as the RTX 2000 Ada, come both as an embedded MXM module and as a low-profile card.
-
//
System-on-module GPUs. NVIDIA Jetson modules combine Arm processor cores, an NVIDIA GPU and AI engines on one module running JetPack. Our guide to NVIDIA Jetson covers the range.
-
//
Integrated graphics. The GPU built into an embedded x86 or Arm processor, such as Intel Iris Xe or Arc graphics in embedded Core processors. It shares system memory and adds little power or heat.
Embedded GPU vs consumer GPU
The same chip can be sold both ways. These are the differences that matter when a system has to be built and supported for years:
| Embedded GPU | Consumer GPU | |
|---|---|---|
| Availability | Planned for long production runs | Follows the consumer product cycle |
| Power | Fixed ratings, 35 W to 115 W on our MXM modules | Set for peak performance, often far higher |
| Cooling | Heat spreader, fan cooler or conduction to the chassis | The card’s own fans |
| Format | MXM module, low-profile card or system-on-module | Full-height card |
| Memory | ECC on many current modules | Usually no ECC |
| Drivers | Professional drivers for long-lived systems | Frequent game-focused updates |
MXM vs PCIe GPU in an industrial PC
When an x86 industrial PC needs a discrete GPU, the choice is usually between an MXM module on the board and a PCIe graphics card in a slot. Both use the same families of NVIDIA GPU; the trade-off is space, cooling and the upgrade path.
| MXM module | PCIe graphics card | |
|---|---|---|
| Mounting | Flat on the board, screwed down | Upright in an x16 slot |
| Cooling | Heat spreader into the chassis, or a fan cooler | The card’s own fan, or system airflow for passive cards |
| Power | 35 W to 115 W on our modules | Up to what the slot, power supply and airflow allow |
| Sealing and vibration | Suits sealed, fanless, vehicle and rail systems | Needs airflow and a card retainer under vibration |
| Upgrades | Same type and power class, checked as a system change | Swap the card within slot power and space |
| Suits | Compact, sealed and mobile systems | Cabinets, racks and box PCs with room for a card |
In practice, choose MXM when the enclosure is sealed, small or on a moving vehicle, and a PCIe card when the system sits in a cabinet or rack with room for airflow and the largest GPU it can take. Our MXM GPU computers cover the first case, and our industrial GPU computers include box PCs with PCIe slots for the second. Some rugged box PCs hold a PCIe card in a sealed expansion bay, which sits between the two, and our rugged GPU computers include both formats.
Embedded GPU or Jetson
Jetson modules are often the first embedded GPU people meet, and for many vision and robotics jobs they are the right one, with the GPU, Arm cores and AI engines sharing one module and one power budget. An x86 system with an MXM or PCIe GPU makes more sense when the software is built for x86 and Windows, several models or many camera streams run at once, or the application needs more GPU memory than a module offers. Our guide to AI accelerators compares the other options, including NPUs.
How to choose an embedded GPU
Start with the workload, then work outwards to the system:
-
//
Model and memory. Check that the model, its batch and its frames fit in GPU memory. The MXM GPU modules we supply run from 4 GB to 16 GB.
-
//
Power and heat. The enclosure has to shed the module’s full rating at the highest ambient temperature on site, not just at room temperature.
-
//
Displays. Display builds drive three or four DisplayPort screens; headless builds run compute beside the processor’s own graphics.
-
//
Software. Check driver support for your operating system and the CUDA, TensorRT or other framework versions your software needs.
-
//
Lifecycle. Ask how long the GPU will stay available and how changes will be announced.
Quick answers on embedded GPUs
Embedded GPUs: common questions
Is an embedded GPU the same as integrated graphics?
Not always. Integrated graphics is one type of embedded GPU, but the term also covers discrete MXM modules and embedded cards with their own memory.
Does NVIDIA make embedded GPUs?
Yes. NVIDIA’s RTX Embedded GPUs are built by module makers into MXM modules and other formats, alongside its Jetson modules.
Can an embedded GPU run AI?
Yes. Current NVIDIA embedded GPUs have Tensor Cores and run CUDA and TensorRT, the same software used on larger GPUs.
Do embedded GPUs need fans?
Not always. Many rugged systems cool the module through a heat spreader into the chassis, while higher-power modules may need airflow.
We fit the GPU, cooling, drivers and CUDA software to suit the workload and the enclosure as standard on every GPU computer we supply.