Vapor Chambers in Blade Server CPU Cooling Solutions
Current blade processors need air cooling solutions that dissipate 100-300 watts with heat sinks that are less than 30 mm high. In order to cool these processors, the heat sink base has to grow in length and width to compensate for the lack of available height. As these dimensions grow, decreasing the base spreading of the heat sink becomes an important factor is decreasing the overall resistance of the heat sink. A vapor chamber used as a substitute to common copper and aluminum as the base of these heat sinks can increase performance by 20-25%. A vapor chamber is a wicked two phase heat transport system that significantly reduces the spreading resistance in applications where there is high heat flux processors coupled with large heat sink areas. In this paper, a CFD model will be constructed to predict the performance gains realized by using a vapor chamber base in lieu of copper or aluminum bases. These predictions will then be experimentally tested to confirm the modeling parameters and the actual improvement of the heat sink. By utilizing vapor chambers in the heat sink design, thermal engineers will gain valuable heat sink performance within the constraints imposed by the blade system architecture.
Blade Server Growth
Industry trends are moving towards getting as much computing power into the smallest spatial footprint. Blade servers are able to condense the performance of a multiple 1U racked computer system into half the size. This data from IMEX Research suggests the revenue of blade servers increase significantly year over year.
Blade Servers optimize the space limitations by essentially putting all of the computer processing power on hot swappable “blades” that are usually about less than 40 mm tall. This severely restricts the amount of surface area above the CPUs for cooling of the processors.
The heat sink utilized in this analysis was 134 mm inches in the flow length, 110 inches perpendicular to flow and a total of 14 mm tall. These dimensions should be should be comparable to most current blade cooling heat sinks. The fin was 10 mm tall, 0.2 mm thick copper with 2.3 mm fin pitch. Both the metallic heat sink and the Thermabase was 4 mm thick. The fins were “interlocked” and attached to the base via lead-free solder.
The CPU was simulated with a 20 mm x 20 mm heat source expelling 130 watts of heat. The airflow was fully ducted ranging from 8 to 23 CFM.
A vapor chamber is essentially a flat heat pipe that can spread heat evenly in two dimensions. It consists of a hollow copper shell that has an internal wick structure to transport the fluid to the heat source and a vapor area for the transport of the vapor away from the heat source. The entire chamber is operated under vacuum so that the operation fluid (water) vaporizes at room temperature. This is how the vapor chamber very efficiently transports heat evenly across the surface of the condenser, allowing all the fins to work as hard as the ones that are over the center of the heat source. The vapor chamber performance will increase over that of a metallic heat sink the larger the base area and the smaller in the heat input.
Modeling the vapor chamber can be done by breaking the part down into four “sandwiched” cuboids. The top and bottom wall of the vapor chamber will be copper with a conductivity of 380 watts/mK. The thickness of the wall is dependent on the necessary strength of the overall chamber and the amount of spreading needed with in the copper walls prior to reaching the wick. The wick conductivity ranges from 30 to 60 watts/mK and is dependent on the flux of the heat source. The thickness is more dependent on power and orientation. Lastly the vapor space is the high conductivity feature of the model. This value can range between 5000 to 50000 Watts/mK, depending mostly on the length of the heat pipe or vapor chamber that you are modeling. Ideally there should be about two to three degrees ΔT in the vapor portion of the heat pipe or vapor chamber.
In order to evaluate the performance gain by switching to a vapor chamber, four different models were created. For both the vapor chamber and the metallic base, the simulation was run for the chip being centered on the heat sink and a case where the CPU was off center by roughly 20mm (perpendicular to the airflow). The Thermabase solution shows a 21% increase in performance over the copper base of the same thickness when the CPU is centered on the heat sink. The gain increases to 27% when the CPU is offset from the center of the heat source by 20 mm (in the direction perpendicular to the airflow).
As shown in the figures above, the spreading of the copper base is relatively poor. The gradient from the hottest point of the base to the coolest is about 26°C in the centered CPU model and about 31°C in the off centered model. For the Thermabase, this gradient is much less allowing all the fins to dissipate heat at the same rate.
Since the airflow in a Blade system is primarily dependent on the pressure drop of the heat sink, a Command Center was setup to model the heat sink performance at a range of airflows (8 to 23 CFM in increments of 3 CFM). Although the pressure drop is not shown as an output variable in the set of airflows in this model, this added feature can be shown in the future to better understand the effect of airflow on the pressure drop of the heat sink.
To validate the results of the FloTHERM model, a bench top cooling system was built to simulate the modeled heat transfer. This experimental unit consisted of the heat sink, a calibrated flow bench, a duct sized to completely enclose the heat sink (minimal bypass), and removable heater block with thermocouple. The flow bench was an Airflow 2000 that measured the flow via curves that relate the internal air flow to the pressure drop across a sized orifice. The heater block itself was made of copper and spring loaded to ensure flat interface with the heat sink (thermal grease was used and the TIM). At the center of the heat source was a spring loaded thermocouple that’s continuity was complete only when the heat sink was loaded into the heater block. This ensures reading was done on the hottest section of the base and not within the heater block itself. Another thermocouple was located in the upstream flow of the duct to measure ambient. The spring-loaded thermocouple in the base minus the ambient temperature divided by the over power resulted in the resistance reported.
FloTHERM modeling of the heat sink was very accurate by estimating within 10% of the actual experimental value. In both the 2-phase vapor chamber and the metallic heat sink the prediction values were slightly higher than the tested values, skewing towards more conservative numbers which is ideal in most modeling scenarios. More importantly the overall predicted improvement (% gain) of switching to a vapor chamber was very accurate when compare to experimental data (within 5% in the instance).
By changing the copper base to a Thermabase the overall heat sink performance heat increased by 21% (27% off-centered CPU) without changing the form or fit of the sink. This offers thermal engineers a unique opportunity to design improvements into their systems without having to change electrical architecture or chassis size.
< Back to Articles