Effective Fan Strategies
Faced with the thermally unfriendly environment of the server chassis, many mass-market graphics processors cards contain an active heat sink, incorporating a small fan within the cooling fins. For embedded XMC modules, a cutout opening in the PCIe adapter card directly under the hottest components allows clearance for an aluminum heat sink that conducts heat through the cutout to fins on the other side of the adapter card. An embedded fan forces air across the fins for active cooling.
This strategy works extremely well for PMC/XMC module adapters for high-power FPGA-based embedded system boards like the one shown in Figure 3. This adapter reduces FPGA silicon junction temperatures by 10C or more.
Air flow sensor probes can be helpful to locate the dead spots and determine if moving a fan or using a different fan with a higher CFM rating could be more effective. Since air flow is hard to visualize, the best solution is often found by trial and error.
Figure 3. An active-cooler heat sink, attached through a cutout hole in the PCIe adapter card, effectively removes heat from the hot components on the inside surface of the attached XMC.
Figure 4. A top-mounted 120-mm fan blows air down onto the expansion cards.
One or more case fans placed directly above the expansion cards blowing downward between the cards is one of the most effective cooling techniques. Since this type of fan is not provided in most server chassis, a special bracket or mounting plate must be fabricated that attaches to the chassis. This requires a chassis with a height of at least 4U to provide vertical clearance for the fan itself and a large enough gap under the top cover to allow adequate air intake.
Figure 4 shows a single 120-mm case fan mounted on a custom plate in a 4U rack-mount chassis. When properly implemented, this type of fan can provide up to 15C of additional cooling.
Another strategy for optimal cooling is simply moving the expansion cards to slot positions that offer the best airflow or clearance. Often, the best slot is the end slot toward the center of the chassis. Although a server motherboard typically has a maximum of seven card slots, the choice of slots is often limited by other considerations. Each one can differ in its interface (PCI or PCIe) and in the number of PCIe lanes. The number of PCIe lanes defines the maximum potential data transfer bandwidth of a card slot, so cards requiring the highest transfer rates are only supported by certain slots.
In addition, the PCIe chipset often imposes certain maximum lane restrictions on combinations of slots used. For example, a “by-16” slot could be reduced to "by-8” if another card is plugged into the adjacent slot. So, the system designer must often trade off between the card slot positions and the data transfer rates to achieve the best cooling configuration.
An Overall Approach
Successful thermal management requires multiple strategies, both at the board design level and during system integration. Embedded cards should be equipped with thermal sensors and software monitoring utilities. Taking advantage of these sensors plus sensors for motherboard components and disk drives can help during system integration and testing, and also benefit the end user.
Conducting heat from hot components like FPGAs through metal structures to fins that are exposed to airflow is essential for maintaining silicon junction temperature margins. Fans incorporated within the card heat sinks, or mounted separately, provide targeted airflow in the poorly ventilated card slot area of most server chassis. Finally, proper card-slot relocation can make a major difference.
Experience is invaluable in the often elusive and frustrating task of cooling embedded PC systems, but applying these common sense strategies will help system designers find the best solution.
Rodger Hosking is vice-president and co-founder of Pentek, where he’s responsible for new product definition, technology development, and strategic alliances.