Browse Prior Art Database

Auxiliary Cooling for High-Power and GPU Accelerator Adapters in a 4U Server Disclosure Number: IPCOM000246443D
Publication Date: 2016-Jun-07
Document File: 4 page(s) / 276K

Publishing Venue

The Prior Art Database


Disclosed is a method for auxiliary cooling for high-power and graphics processing unit (GPU) in a rack unit.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 42% of the total text.

Page 01 of 4

Auxiliary Cooling for High -

-Power and GPU Accelerator Adapters in a

Power and GPU Accelerator Adapters in a 444U Server

U Server

Disclosed is a method for auxiliary cooling for high-power and graphics processing unit (GPU) in a rack unit. Heat is the number one enemy to over-clocking; therefore, PC components must have sufficient cooling. Keeping the CPU, GPU, and other system components cool is vital to ensuring a successful operation. Without proper cooling, higher temperatures lead to system instability and unpredictable behavior. Theoretically, frequency is directly proportional to power and temperature. The higher the frequency, the more power consumed and the hotter the device becomes. In order to achieve higher frequencies, component temperatures should be kept as low as possible. The most commonly used cooling techniques are air cooling with fan mounted or passing heat sinks and water cooling blocks. As water cooling is harder to install and generally much more expensive than air cooling for server conversion, passive air cooling technological advancements have made air cooling a very practical option for GPU accelerators in server PCIe formats. The disclosed method has the following features:

Overall unique method of cooling both the CPU and GPU accelerator to maximum



Configurable, only 1 fan per GPU accelerator adapter required. Applies to both socket and


2-socket servers without additional base system air movers. Not associated with base system cost structure.

Use of counter-rotating (CR) fan provides cooling redundancy at the GPU accelerator level,


not just system level redundancy. Adapters with their own active cooling are rarely redundant.

Minimizes acoustic level increases.


Provides added DASD cooling.


Utilizes separate, unique airflow inlet.


    Requirements for a specific graphics adapter in a 4U server are used to describe an example application of the disclosed method. The cooling specifications for passively cooling the Nvidia* Tesla* K80 GPU accelerator, double-wide adapter requires a specific airflow rate as a function of inlet temperatures. In some electronic server enclosures, the only PCIe x16 slots able to accommodate such adapters are only located directly behind the server's CPU. These processors can be ran anywhere between 190W and 300W. For maximum system performance, running at the highest processor power is desired. For the Nvidia Tesla K80 dissipating 300W, the temperature of the airflow entering the GPU accelerator adapter is a function of airflow bulk air temperature versus volumetric rate. Figure 1 depicts a general correlation.

Figure 1

The required airflow for a given air temperature entering the GPU accelerator in a 25C


Page 02 of 4

room is shown in Figure 2. At 40C inlet, the flow rate required is 23.8 CFM; at 50C inlet, it increases to 31.4 CFM. If there were a 300W CPU upstream of the GPU accelerator adapter for a 24C inlet airflow, the required airflow rate...