How to Select a Fan and Heat Sink for NVIDIA Jetson, Intel Core Ultra, and Other AI Inference Hardware

AI inference hardware - NVIDIA Jetson modules, Intel Core Ultra-based embedded systems, Hailo and Rockchip NPU platforms, and similar compute-on-module designs - packs serious processing power into tight footprints. The thermal demands vary by platform, but the selection challenge is consistent: the wrong fan-and-heat-sink combination will cost you performance, reliability, or both in a production environment.

Selecting the right combination is not difficult if you approach it systematically. This post walks through the process end-to-end - how to build the thermal budget, how to choose a heat sink geometry, how to select a fan, and how to match the two for the environment the hardware actually operates in.

Step 1: Build the Thermal Budget Before You Select Anything

Fan and heat sink selection starts with numbers, not with browsing a catalog.

The inputs you need:

• TDP of the processor - the maximum sustained heat dissipation you need to manage; this ranges from 5W for lower-power NPUs like the Hailo-8L to 60W+ for high-performance platforms like the NVIDIA Jetson AGX Orin or Intel Core Ultra in sustained inference workloads
• Maximum junction temperature - specified by the chip manufacturer, typically 85-105°C
• Target junction temperature under load - design to 70-80% of maximum junction temperature for reliable long-term operation
• Maximum ambient temperature - not your lab temperature; the temperature inside the enclosure in the deployment environment

The thermal resistance budget is: R(total) = (T(junction) - T(ambient)) / TDP

That total thermal resistance is distributed among the thermal interface material, heat sink, and airflow. Work backward from your budget to understand what each element needs to deliver.

Step 2: Heat Sink Selection

For forced-air cooling of AI inference hardware, the most common heat sink geometries are:

Pin fin heat sinks - good omnidirectional airflow performance; well-suited to applications where airflow direction is not tightly controlled or where a blower provides radial airflow.

Straight fin heat sinks - optimized for axial fan configurations with controlled airflow direction; lower static pressure drop than pin fin in the intended orientation, which means you can use a less powerful fan to achieve the same thermal performance.

Skived or folded fin heat sinks - higher fin density than extruded profiles, better performance per unit volume; used when footprint is constrained, and performance demands are high.

Heat sink sizing comes down to thermal resistance (°C/W) at a given airflow rate. Match that to your thermal budget at the airflow your fan will actually deliver, given the static pressure the heat sink imposes.

Step 3: Fan Selection

For AI inference enclosures, you are typically choosing between:

Axial fans - move high air volume at low static pressure; best for open or low-restriction airflow paths; common sizes for AI inference applications are 40mm, 60mm, and 80mm; smaller fans run at higher RPM to maintain airflow, which means more noise and more bearing wear.

Blowers (centrifugal fans) - move air at higher static pressure, suited to confined channels and restricted flow paths

Key fan parameters to match your application:

• Air volume at operating static pressure - use the fan curve, not the free-air spec; the static pressure imposed by your heat sink fins and enclosure routing determines where you operate on that curve
• Noise (dBA) - a larger, slower-spinning fan is almost always quieter than a smaller fan running fast to achieve the same airflow
• Operating temperature range - ensure the fan is rated for the internal ambient of your enclosure, not the external ambient
• Bearing type and MTBF - for industrial or long-life applications, ball bearing fans provide better MTBF at elevated temperatures
• Speed control - PWM control is standard for variable-speed fans in embedded applications; match the control signal to your processor's fan header or fan controller IC; most Jetson platforms and Intel Core Ultra designs expose a PWM fan header directly on the carrier board

Step 4: Match Fan and Heat Sink as a System

The most common mistake here is selecting a heat sink and a fan independently and assuming they will work together. Fan performance is described by a pressure-flow curve. Heat sink performance is described by thermal resistance as a function of airflow. You need to find the operating point where the fan's output matches the heat sink's static pressure requirement - that is, where you read the actual airflow, and from that the actual thermal resistance.

If the operating point does not hit your thermal budget, your options are a higher-performance fan, a lower-restriction heat sink, or both.

Work with an Expert Supplier

YS Tech USA stocks fans across the full range of sizes relevant to AI inference hardware - from 25mm to 172mm in DC configurations, with EC blower options for higher-static-pressure applications - alongside extruded and custom heat sink profiles. If your application is standard, you can specify from our product range. If you are working through a non-standard form factor, a tight thermal budget, or an enclosure design that does not match a catalog solution, our engineering team works directly with OEM engineers on selection and custom design.