Design Considerations for Cooling High Heat Flux IC Chips With Microchannels Satish G. Kandlikar Rochester Institute of Technology Editor s notes: Thermal emergency in integrated circuits has become an important issue with aggressive scaling trends. Several novel cooling techniques are investigated in both academia and industry. Sophisticated active cooling techniques are required to mitigate the thermal issues faced by the chips in the current and future technologies. VPartha Pande, Washington State University h TRADITIONAL IC CHIP cooling techniques involve using a heat spreader to dissipate heat over a larger heat transfer surface area cooled by air or liquid. The heat spreader itself introduces conduction resistance, which adds to the temperature drop between the chip surface and the surfaces dissipating heat to the coolant. Direct cooling of chip in a 2-D architecture eliminates this resistance and is being actively considered for heat fluxes of 70 100 W/cm 2. For example, introduction of a coolant chip with embedded liquid cooled microchannels on the back of the chip provides a viable cooling solution Digital Object Identifier 10.1109/MDAT.2014.2299535 Date of publication: 13 January 2014; date of current version: 07 August 2014. in the IBM chip cooler design by Colgan et al. [1] for heat fluxes approaching 1000 W/cm 2. The 2-D multicore chips enabled by modern silicon technologies dissipate more than 100 W over very small areas. With the advent of 3-D technologies, the power dissipation densities as well as the heat flow path lengths are increasing. Consequently, conventional cooling techniques are no longer effective for these high-end systems-on-chip (SoCs) and multicore chips. The microchannel cooling with liquid is particularly suitable in a 3-D architecture as the constrained height requirement for the channels makes air cooling or other options such as jet impingement and spray cooling not feasible. Limiting the individual chip heat fluxes in a 3-D stack below 100 W/cm 2 enables implementation of current 2-D liquid cooling designs with microchannnels. For example, the network-on-chip (NoC)-based multicore architecture proposed by Vangal et al. [2] at INTEL utilizes 80 tiles arranged as 8 10 2-D arrays of floating point cores and packet-switched routers operating at 4.27 GHz with a footprint area of 275 mm 2 and a July/August 2014 Copublished by the IEEE CEDA, IEEE CASS, IEEE SSCS, and TTTC 2168-2356/14 B 2014 IEEE 43
Tutorial total heat generation rate of 97 W. This NoC uses layered protocols and on-chip routers, links, and well-defined interfaces and can be used as the basic building block of a network tile [2] in a 3-D architecture. Microchannel cooling Channel classification It is desirable to bring the coolant directly to the chip surface to eliminate the conduction resistance in a heat spreader. The channel dimensions for the coolant flow are therefore smaller than in a conventional compact heat exchanger as they are limited by the maximum thickness of the silicon substrate. The channels are characterized by their hydraulic diameter which is defined as Table 1 Channel classification scheme [3]. D h ¼ 4 A c (1) P w where A c is the cross-sectional area, and P w is the wetted perimeter of the channel cross section. Thus, for a circular channel the hydraulic diameter is the same as the channel diameter, while for a rectangular channel with sides a and b, the hydraulic diameter is 2ab=ða þ bþ. The concept of hydraulic diameter is useful as it allows heat transfer or pressure drop correlations developed for the commonly used circular geometry to be applied to channels of other cross sections. As the channel size becomes smaller, the scale effects begin to affect the flow. In single-phase gaseous flow at atmospheric pressure, the mean free path and rarefaction effects become relevant at around 200-m channel size, while in two-phase flow, the capillary forces become important at these dimensions. However, for liquid flow the flow remains unaffected from molecular considerations down to a few micrometers. The classification of channels given in Table 1 is based on the manufacturing and gas/liquid/twophase flow considerations and is frequently employed in defining microchannels and minichannels. The heat transfer performance is improved for smaller channels, but the pressure drop becomes unacceptably high. As a result, channel dimensions in the neighborhood of 200 m are considered as optimal in the IC chip cooling application [4]. Smaller channels result in higher pumping power, and also the high pressure in the microchannels causes large mechanical stresses in the silicon channel walls. Fabrication of these channels can be accomplished with the standard microfabrication techniques, such as deep reactive-ion etching (DRIE) or other suitable processes. Before introducing the heat transfer and pressure drop equations for microchannel flows, a brief background of fluid flow and associated heat transfer is presented next. Laminar and turbulent flow A flow can be classified as laminar or turbulentv an indication of the level of eddy mixing within a flow. A laminar flow has parallel fluid streamlines and the heat transfer process is less efficient as it is governed mainly by diffusion (heat conduction). On the other hand, turbulent flow consists of local eddies which lead to a more efficient heat transfer process. The flow is characterized by the Reynolds number, which is a nondimensional parameter quantifying the flow structure. It is defined as Re ¼ VD h (2) where is the fluid density (kg/m 3 ), V is the mean fluid velocity (m/s), and is the fluid viscosity (N-s/m 2 ). For a circular channel, the flow transitions from laminar to turbulent as Re increases beyond 2300. The transition Re is slightly different for other channel geometries. For microchannels employed in cooling application, the flow is generally laminar because of the small hydraulic diameter. Single-phase liquid heat transfer in microchannels Since the flow is laminar, a simple relationship holds for predicting the fully developed heat transfer coefficient h, which is defined as the heat transfer 44 IEEE Design and Test
rate per unit area per unit temperature difference (W/m 2 - C) between the local heated surface T w and the bulk fluid T f h ¼ q 00 ðt w T f Þ (3) where q 00 is the heat flux at the channel wall (W/m 2 ). The heat transfer coefficient is often presented in a nondimensional form using the Nusselt number Nu, defined as Nu ¼ hd h (4) k where k is the thermal conductivity of the fluid (W/m- C). For the fully developed laminar flow, the Nusselt number has a constant value, which depends on the mode of heating, either constant heat flux (commonly used in a direct chip cooling system design) or constant surface temperature (usually applicable in a cold plate design). The Nusselt number also depends on the channel geometry. For rectangular channels, it further depends on the channel aspect ratio defined as width/height. With a constant heat flux boundary condition, Nu is 4.36 for a circular cross section while for a square cross section, Nu is 3.61. The heat transfer coefficient however, is significantly higher near the channel entrance due to the entrance effects arising from developing velocity and temperature profiles. The specific equations can be found in [5]. A well-designed cooling system removes the heat while keeping the local wall temperature below a desired upper limit. Another equation that is needed in the cooling system design relates the total heat transfer rate with the temperature rise and the mass flow rate of the liquid coolant Q ¼ mc p ðt f ;out T f ;in Þ (5) where Q is the total heat transfer rate in the heat exchanger (W), m is the liquid flow rate (kg/s), C p is the specific heat of liquid (J/kg- C), and T f ;in and T f ;out are the liquid temperatures at the inlet and outlet sections ( C). The coolant temperature thus increases as it flows through the heat exchanger, and its cooling ability is accordingly reduced for a given maximum allowable wall temperature. A higher flow rate will reduce the temperature rise for a given heat input but it will increase pressure drop and pumping power, as described in the next section. Single-phase liquid pressure drop in microchannels As liquid flows through the microchannels, the pressure decreases from the inlet section to the outlet section due to frictional losses. This pressure drop p needs to be overcome by an external pump circulating liquid coolant. The pumping power to overcome the flow resistance is given by P ¼ m p (7) where P is the pumping power (W), m is the coolant mass flow rate (kg/s), and is the density of the coolant (kg/m 3 ). A commonly employed metric for performance evaluation is the coefficient of performance (COP). It is the ratio of the heat removed in the heat exchanger to the pumping power expended in circulating the coolant through the channels COP ¼ Q P : (8) A COP value above 30 50 is desirable in cold plates. When the cooling requirements are less stringent, such as higher allowable surface temperatures, it may be possible to attain a higher value of COP.ACOPofatleast 20 is desirable for chip cooling with microchannels. The microchannel geometry can be optimized by changing the channel and fin dimensions, along with the flow rate, to meet a specific design requirement. An overview of the optimization methodology and guidelines for designing plain microchannels to cool IC chips is provided by Kandlikar [6]. Although the microchannels offer a high thermal performance, further improvements can be obtained by employing different enhancement techniques to provide a high heat transfer coefficient with low pumping power. Offset strip fin geometry is one of the most effective geometries in enhancing heat transfer, and its use with the variable density fin pattern presented limits the pressure drop, as discussed in the next section. Enhanced microchannels Offset strip fins A number of heat transfer enhancement techniques have been investigated in literature, including fins, dimples, flow bifurcations, wavy fins, etc., for example, [7] and [8]. The heat transfer coefficient is July/August 2014 45
Tutorial Figure 1. IBM microcooler with offset strip fins and multiple inlet/outlet ports in a 10-mm square IC chip [1]. Reproduced with permission from the IEEE. significantly higher for short fins as new boundary layers are formed over their surfaces. This fact is utilized in designing the offset strip-fin configuration. A practical implementation of this geometry by Colgan et al. [1] at IBM is shown in Figure 1. The straight walls of the microchannels are replaced with 200 500-m-long fins. The width of the flow channel between the fins is also in the same range of 200 500 m. Subsequent rows of fins are offset by half the channel width so as to provide the offset strip-fin configuration. The heat transfer coefficient in this IBM microcooler is several times higher than a plain microchannel, depending on the fin length. To keep the pressure drop reasonable, the flow length of over a 10-mm-long chip is reduced to only 3 mm by providing multiple entry and exit ports. The heat transfer coefficients in these geometries were reported to be over 500 000 W/m C with COPs rangingbetween40and133[9].thedesignisquite attractive, but requires complex manifold cover to provide the short flow lengths. Pin fins Pin fins of different cross sections, such as circular, elliptical, and tear drop shapes are placed in the microchannel flow passages or in the flow field, similar to the offset strip fins shown in Figure 1. The pin fins further enhance the heat transfer, but are associated with steep increases in pressure drop when compared to plain microchannels. Peles et al. [10] placed circular pin fins, 100 m in diameter and 243 m tall, at both transverse and longitudinal pitches of 1.5 mm in a 1.8-mm 10-mmlong silicon channel. They could dissipate a heat flux of 790 W/cm 2 with a pressure drop of 100 kpa and a water temperature rise of 30.7 C. The resulting COP is 642, which is very attractive. However, the pressure drop of 100 kpa (with an inlet pressure of at least 200 kpa to avoid vacuum in the coolant system) is too high for practical implementation in IC chips. Kosar and Peles [11] investigated the effects of different fin shapes on heat transfer and pressure drop performance. They compared several circular, one hydrofoil-shaped, one cone-shaped (with semicircular leading edge), and one rectangular cross-section fins. They concluded that the best performing geometry depends on the flow rate and allowable pressure drop. In general, unstreamlined (staggered) circular pin fins were found to be desirable if high pressure drops are acceptable. The COPs for these geometries are similar to those in their earlier study [10]. The high pressure drops in the pin fin geometries, in excess of 100 kpa, are above the allowable limit in practical IC chip cooling application. These geometries may be considered in Colgan et al. s [1] microcooler configuration with shortened flow lengths. These pin fins however are quite attractive within the coolant channels of copper cold plates, which can withstand higher pressures. Variable fin density for uniform temperature Temperature uniformity of the IC chip is another consideration in designing a cooling system. Large temperature variations in the devices placed on the chip along the coolant flow direction are undesirable as they introduce thermal stresses as well as performance concerns. The coolant receives heat as it flows through the channel, and its temperature rises from the inlet section to the outlet section as given by (5). This temperature rise is related to the heat removed from the chip and the coolant mass flow rate. A lower flow rate is desirable to keep the pressure drop low, whereas a higher flow rate is desired to limit the temperature rise of the coolant. The chip surface temperature T w at any location is related to the local fluid temperature T f and the local heat transfer coefficient h by T w ¼ T f þ q00 h : (9) For an enhanced microchannel with a dense array of fins, the heat transfer coefficient will be high 46 IEEE Design and Test
and almost constant along the flow path, and the chip surface temperature will rise correspondingly with the coolant temperature. Hence the coolant and chip surface temperatures will attain their maximum values at the outlet section. This maximum chip surface temperature occurring at the exit should be kept below the maximum allowable value for safe operation. Assuming a constant heat transfer coefficient, the chip surface will be cooler at the inlet due to the cooler fluid temperature at the inlet. A lower heat transfer coefficient may thus be acceptable near the inlet section in an effort to keep the chip temperature here closer to that at the outlet. Rubio- Jimenez et al. [12] proposed a variable fin density configuration in which the fin density decreases from the inlet to the outlet, thereby gradually increasing the heat transfer coefficient and the surface area along the flow direction. In a subsequent paper, Rubio-Jimenez et al. [13] compared the cooling performance of two 10-mm 10-mm square chips with inline and offset strip fins arranged in variable fin density configurations shown in Figure 2. Each configuration consists of three zones with different fin densities. The pressure drop was lower than the plain microchannels. They attained the same pressure drop as Colgan et al. s design [1] with shorter flow length of 3 mm. Moreover, the temperature variation over the surface was reduced to less than 5 C. The COP of the variable density fins will be higher than the offset strip fin design, increasing inversely in proportion to the pressure drop reduction. Since the fabrication cost of the variable density fins would be the same as the uniform density fins, their benefits from savings in reduced pumping power and temperature uniformity make them attractive Figure 2. Rectangular fins with variable fin density along the flow length in inline and offset fin configurations [12]. Reproduced with permission from the IEEE. alternatives for plain microchannels or microchannels with uniform density fins. Structured roughness enhanced microchannels Introducing structured roughness in the form of longitudinal ribs on the channel walls is another technique that provides significant heat transfer enhancement with relatively low pressure drop penalty. Lin and Kandlikar [14] investigated minichannels with a sinusoidal roughness profile on the channel walls as shown in Figure 3. They found that for a roughness element height to channel hydraulic diameter ratio of 11.1, the transition to turbulence occurred at a relatively low Reynolds number of 1000, and the heat transfer coefficient was July/August 2014 47
Tutorial Figure 3. Structured roughness with sinusoidal fin profile on the channel walls [14]. Reproduced with permission from the American Society of Mechanical Engineers (ASME). enhanced by a factor of 3.9. Although the pressure drop also increased, it was found that the ratio of heat transfer enhancement to friction factor enhancement (directly related to pressure drop) was the highest for this structured roughness as compared to other enhancement techniques, including twisted tapes and coiled inserts reported in the literature. The main reason for this favorable result is the proximity of the enhancement features to the heat transfer surface, and the smooth profile at the ridges that does not introduce significant disturbance in the core flow field. This technique could be further modified to include variable pitch or variable fin height in the flow direction to utilize the available pressure drop more effectively and provide temperature uniformity, similar to the variable density fins discussed in the previous section. Two-phase cooling with microchannels Flow boiling in microchannels is a very attractive technique because of its two main benefits: 1) relatively uniform wall temperature during the boiling process; and 2) a low fluid circulation rate due to large latent heat of vaporization of the working fluid. Currently, intense research is being conducted on providing stable operation and improved heat transfer performance [15]. It is expected that in the near future this technique will be explored further and will be ready for practical implementation in IC chip and cold plate cooling applications. Comparison of different cooling options Each of the available microchannel-based IC chip cooling options offers specific advantages in terms of heat transfer enhancement, pressure drop requirements, and temperature uniformity. Table 2 provides a comprehensive summary of their relative performance in these three categories. Microchannels offer a very promising heat removal technique for cooling 2-D and 3-D IC chips. They can be directly fabricated on the back of an 48 IEEE Design and Test
Table 2 Evaluation of different microchannel-based IC chip cooling options. IC chip, or in a cooling chip bonded to the IC chip. Plain microchannels offer a very high coefficient of performance (COP ¼ heat removed/pumping power) and are recommended for heat fluxes up to 70 W/cm 2. For higher heat fluxes, enhanced microchannels are recommended. In particular, offset strip fins with variable fin density configuration provide the most effective utilization of the available pressure drop. Future Research Needs THE VARIABLE FIN density concept needs to be further investigated through optimization of the fin cross section and the fin density distribution. Structured roughness features are also seen as a very promising technique. They have been studied in literature for minichannels with channel hydraulic diameters of 1 2 mm. It is recommended that they be explored for microchannel applications, and appropriate roughness structure geometry, roughnessheight-to-pitch ratio, and flow rate ranges be identified through an optimization study. Two-phase cooling is a very attractive technique, and further research is warranted in developing practical systems that offer stable operation with high heat transfer performance. Finally, these techniques should be evaluated carefully for 3-D IC cooling applications where the pressure drop and space constraints are even more stringent, with added limitations imposed by the placement of the through silicon vias (TSVs). h July/August 2014 49
Tutorial h References [1] E. G. Colgan et al. A practical implementation of silicon microchannel coolers for high power chips, in Proc. 21st Semicond. Thermal Meas. Manage. Symp., San Jose, CA, USA, Mar. 15 17, 2005, DOI: 10.1109/STHERM.2005.1412151. [2] S. R. Vangal et al. An 80-tile sub-100-w TeraFLOPS processor in 65-nm CMOS, IEEE J. Solid-State Circuits, vol. 43, no. 1, pp. 29 41, Jan. 2008. [3] S. G. Kandlikar and A. J. Grande, Evolution of microchannel flow passagesvthermohydraulic performance and fabrication technology, Heat Transfer Eng., vol. 24, no. 1, pp. 3 17, 2003. [4] S. G. Kandlikar and H. R. Upadhye, Extending the heat flux limit with enhanced microchannels in direct single-phase cooling of computer chips, in Proc. 21st IEEE Semicond. Thermal Meas. Manage. Symp., 2005, pp. 8 15. [5]S.G.Kandlikar,S.Garimella,D.Li,S.Colin,and M. R. King, Heat Transfer and Fluid Flow in Microchannels and Minichannels, 2nd ed. Amsterdam, The Netherlands: Elsevier, 2013. [6] S. G. Kandlikar, High flux heat removal with microchannelsva roadmap of challenges and opportunities, Heat Transfer Eng., vol.26,no.8, pp. 5 14, 2005. [7]M.S.SteinkeandS.G.Kandlikar, Single-phase heat transfer enhancement techniques in microchannel and minichannel flows, in Proc. 2nd Int. Conf. Microchannels Minichannels, Rochester,NY,USA, pp. 141 148, Jun. 17 19, 2004, DOI: 10.1115/ ICMM2004-2328. [8]G.Xie,S.Li,B.Sunden,W.Zhang,andH.Li, A numerical study of the thermal performance of microchannel heat sinks with multiple length bifurcation in laminar liquid flow, Numer. Heat Transfer A, vol. 65, no. 2, pp. 107 126, 2014. [9]M.S.SteinkeandS.G.Kandlikar, Single-phase liquid heat transfer in plain and enhanced microchannels, in Proc. 4th Int. Conf. Nanochannels Microchannels Minichannels, 2006, pp. 943 951. [10] Y. Peles, A. Kosar, C. Mishra, C.-J. Kuo, and B. Schneider, Forced convective heat transfer across a pin fin micro heat sink, Int. J. Heat Mass Transfer vol. 48, pp. 3615 3627, 2005. [11] A. Kosar and Y. Peles, Micro scale pin fin heat sinksvparametric performance evaluation study, IEEE Trans. Compon. Packag. Technol., vol.30,no.4, pp. 855 865, Dec. 2007. [12] C. A. Rubio-Jimenez, S. G. Kandlikar, and A. Hernandez-Guerrero, Numerical analysis of novel pin fin heat sink with variable fin density, IEEE Trans. Compon. Packag. Manuf. Technol.,vol.2,no.5, pp. 825 833, May 2012. [13] C. A. Rubio-Jimenez, S. G. Kandlikar, and A. Hernandez-Guerrero, Performance of online and offset micro pin-fin heat sinks with variable fin density, IEEE Trans. Compon. Packag. Manuf. Technol. vol. 3, no. 1, pp. 86 93, Jan. 2013. [14] T.-Y. Lin and S. G. Kandlikar, An experimental investigation of structured roughness effect on heat transfer during single-phase liquid flow at microscale, J. Heat Transfer, vol. 34, pp. 101701-1 101701-9, 2012. [15] S. G. Kandlikar, T. Widger, A. Kalani, and V. Mejia, Enhanced flow boiling over open microchannels with uniform and tapered gap manifolds, J. Heat Transfer, vol. 135, pp. 061401-1 061401-9, 2013. Satish G. Kandlikar is a Gleason Professor of Mechanical Engineering at Rochester Institute of Technology, Rochester, NY, USA. His research interests include boiling heat transfer, IC chip cooling, cold plate design, transport processes in microchannels, and fuel cells. Kandlikar has a Ph.D. from the Indian Institute of Technology (IIT) Bombay, Mumbai, India (1975). He is a member of the American Society of Mechanical Engineers (ASME) and the Electrochemical Society (ECS). h Direct questions and comments about this article to Satish G. Kandlikar, Mechanical Engineering Department, Rochester Institute of Technology, Rochester, NY 14623 USA; sgkeme@rit.edu. 50 IEEE Design and Test