Thermal Management of A Multi-core Master Processing Unit (MPU) for An Ultrascalable Computing Platform Ting Cheng1 , Wei Xiong1, Xiaobing Luo13, Suyi Huang1, Zhiyin Gan23, Sheng Liu23* School of Power and Energy Engineering, Huazhong University of Science & Technology, Wuhan, China, 430074 2 School of Mechanical Science and Engineering, Huazhong University of Science & Technology, Wuhan, China, 430074 3 Division of MOEMS, Wuhan National Laboratory for Optoelectronics, Wuhan, China, 430074 *Corresponding Author: Sheng Liu, Email:
[email protected], Telephone: 86-13871251668, Fax: 86-27-87557074 1
Abstract Thermal management of a super computer is one key point in the whole system design. The working temperature of electronic device in the super computer is the important parameter to show the cooling performance. Thermal design was conducted for one super computer. The experiment was conducted to test the temperature of an ultrascalable computing platform in order to evaluate the cooling performance, the results demonstrate that the thermal management by such a simple method is good. Introduction With the increasing clock rate of microprocessors and high density electronic components in super computer[1], power dissipation is becoming a critical problem of super computer design. As we know, electronic devices are especially sensitive of the temperature, their performances descend due to the increasing temperature[2], so thermal issues are becoming especially critical for high-performance super computer. Thermal management of the equipment is becoming a challenging task for designers [3]. In this paper, one simple thermal management method for a super computer was proposed, the experimental study on the thermal management of one MPU was conducted. The experimental test data shows the cooling performance of a pin-fin sink and fan thermal management system is good. Experimental procedures The study goal is to test the cooling system of a multi-core master processing unit (MPU) for an ultra-scalable computing platform. Figure 1 shows the components of this super computer, it consists of several MPU to realize the calculation function. The structure of a MPU is clearly demonstrated in Figure 2, it contains sixteen calculation flashboards as shown in Figure 3, several communication boards and power modules. For all boards and modules, they generate much heat when computer works. To lower the temperature inside the MPU, several fans were distributed inside the MPU as shown in Figure 2. One row of five fans were placed on the cabinet left wall, another row of five fans were placed between the communication boards and the calculation boards in the middle of the whole cabinet. As for each board and module, the heat sink was used for cooling core component. The power of a CPU was nearly 50W when it is full loaded; FPGA has a full power of 30W. To test the cooling effect, thermocouples were placed on the bottom of fin heat sinks.
Figure 1. The picture of a super computer.
Figure 2. The picture of one MPU module.
Figure 3. The picture of one calculation flashboard. Ten thermocouples were attached at ten different heat sinks on calculation flashboards and two thermocouples were attached at two different heat sinks on one communication module. Figure 4 demonstrates the thermocouple distribution on different heat sinks, in Figure 4, six thermocouples were located on the top of CPU heat sinks. They were marked as number 1, 3, 8, 11, 17 and19, six thermocouples were located on the top of FPGA heat sinks. They were marked as number 13, 7, 9, 10, 18 and 20. Figure 5 shows the experimental setup, the temperature data obtained by the thermocouples was transferred to the data acquisition system and reported on
2008 International Conference on Electronic Packaging Technology & High Density Packaging (ICEPT-HDP 2008) 978-1-4244-2740-6/08/$25.00 ©2008 IEEE
1 13 19 8 20 9 3 7 17 18 11 10
54 52 50 48 46 44 42 40 38 36 34 32 30 28 26 24 22 20 18 04:49:39 04:53:51
04:58:03
05:02:20
05:06:34
--
time
Figure 6. Variation of temperature of the CPU heat sinks with the operation time for all of them worked at empty load. 50
1 13 19 20 8 9 18 7 17 11 3 10
48 46 44 42 40
temperature
Result and analysis Since different usage conditions of the super computer will result in different heat load, several cases were tested. The test cases in which the super computer works in three different loads were as follows, 1) all the CPUs were load free; 2) the calculation CPUs worked at between 67% and 69% of full load, the communication CPUs worked at between 73% and 75.5% of full load; 3) the calculation CPUs and the communication CPUs both worked at 99% of full load. The ambient temperature was about 19℃. The results show that when all the CPU load free, the temperature of communication module CPU heat sink was as high as 51℃, all the heat sinks had a short balance time, other temperatures achieved by different thermocouples were shown as Figure 6. When the calculation CPUs worked at between 67% and 69% of full load and the communication CPUs worked at between 73% and 75.5% of full load, the temperature of communication module CPU heat sink was 49 ℃ , all the heat sinks temperature changed little . Other temperatures were showed as Figure 7, when the calculation CPUs and the communication CPUs both worked at 99% of full load, the temperatures of heat sinks were different at first, but several minutes later they balanced at the same degree after the load was free. All these test results demonstrate that the thermal management solution based on the fan and fin could be used.
temperature
the PC monitor, the model of the data acquisition system in the experiment are Keithley 2700 multimeter and control unit 7700.
38 36 34 32 30 28 26 24 22 20 18 02:10:48 02:12:30 02:14:10 02:15:51 02:17:33 02:19:13 02:20:56
time
temperature
Figure 7. Variation of temperature of the CPU heat sinks with the operation time for the calculation CPUs worked at between 67% and 69% of full load, the communication CPUs worked at between 73% and 75.5% of full load.
Figure 4. Thermocouple distribution at various positions of the MPU module.
1 13 19 20 8 9 3 7 17 18 11 10
58 56 54 52 50 48 46 44 42 40 38 36 34 32 30 28 26 24 22 20 18 05:23:39 05:32:08 05:40:3305:48:54 05:57:2206:05:53
time
Figure 5. Experimental setup.
Figure 8. Variation of temperature of the CPU heat sinks with the operation time for the calculation CPUs and the communication CPUs loads change. According to the data shown in Figure 8, there is an obvious working temperature difference for different boards at different loads. The calculation program can maintain the electronic devices running at stabilized output. Different working temperatures are caused by the air flow in the box. The spatial arrangement of the boards also affects the temperatures. The air flow blowing direction from the fans is
2008 International Conference on Electronic Packaging Technology & High Density Packaging (ICEPT-HDP 2008)
stochastic. Air flow in the cabinet is non-form, especially when it is near cabinet. The flow from two rows fans was interacted. In addition, electronic elements in the MPU disturbed the airflow. All the above factors led to different temperatures. Conclusions Powerful fans, plate or fin heat sinks are still effectively used in computer cooling and they are less cost, easy maintenance. In this paper, heat sink with fans were used for thermal management of super computer, experimental tests were also conducted. The temperature distribution at different locations and different loads were achieved. The results demonstrate the present cooling system works well. Acknowledgments The authors would like to thank Shanghai Redneurons Co. Ltd for providing experimental test environment. References 1. Bhave.Ninad, Okamoto.Nicole, “Modeling noncoplanarity effects on thermal performance of computer chips”, ISAPM. 4419927; 3-5 Oct. (2007), pp.47 – 52 2. Brooks.D, Martonosi,.M., “Dynamic thermal management for high-performance micro-processors”, HPCA, (2001), 903261; 19-24 Jan, pp.171 - 182 3. Nakanishi.M, Nakayama.W, Behnia.M, Soodphakdee.D,” A new approach to the design of complex heat transfer systems: notebook-size computer design”, THERM 2002. The Eighth Intersociety Conference, pp.595 - 599
2008 International Conference on Electronic Packaging Technology & High Density Packaging (ICEPT-HDP 2008)
