154 IEEE TRANSACTIONS ON COMPONENTS, HYBRIDS, AND MANUFACTURING TECHNOLOGY, VOL. CHMT-7. NO. 1. MARCH 1984 Narrow Channel Forced Air Heat Sink NORMAN GOLDBERG, A&tract-The heat transfer between the coolant and the heat sink contributes a significant fraction of the thermal resistanceof a package. It has been widely recognized that the heat transfer can be increased by increasing the area of the .heat sink in contact with the coolant, and therefore the use of fins is common. Howevei it has not been generally recognized that for laminar flow in confined channels, the heat transfer coefficient is inversely proportional to the width of the channel. Narrow channel forced air heat sinks have been designed, built, and tested. Three different channel widths, 5,10, and 25 mils, were investigated. The design of the sinks permit integrated circuits @C’s) on l/2 in centers and printed circuit (PC) boards on 1 in centers. The measured values of individual heat sinks with 30 Vmin air flow are 5.5,3.7, and 3.4”C/W for 25,10, and i mil channel widths, respectively. These values are in good agreement with calculated values. A 16 chip package with chips on l/2 in centers was assembled utilizing 10 mil channel width heat sinks. A plenum fitted over the sinks supplied air to each of the sinks from a single source. The measured values of thermal resistanceof the four center dies were between 5.3 and 5.9”C/W. If the chips are mounted directly on the sink instead of on the alumina substrate, the thermal resistance would be reduced by about 2”C/W. SENIOR MEMBER, IEEE over ambient of the cooled device per unit of power dissipated. In our package the inte:grated circuit is mounted face up on a ceramic subs&ate, and the hkat sink is mounted at the: sam,e position on the other side of the substrate as shown in Fig. 1.l.t is tionvenient to consider the thermal resistance of the heat sinks as three resistances in series as described by Tuckerman and Pease [3] : econd,the resistance due to thermal conduction from the circuits through the silicon, the silicon to substrate bond, the ceramic substrate, ,the substrate to heat sink bond, and the lead-in to the sink; 8,,,,,, the convective resistance, primarily determined by the heat transfer coefficient between the heat sink and the coolant fluid; and oheat,the resistance due to the heating of the fluid as it absorbs energy while flowing through the sink. econdis given by t 8 =- ‘Ond kA INTRODUCTION HE AVAILABILITY of high performance very dense very large-scale integration (VLSI) will result in significant improvement in the performance of computer systems. However, in order to fully utilize VLSI, greatly improved heat removal will be required. In the near future, we must be able to package a 200 mil x 200 mil size chip that dissipates 10 W. It is a challenge to cool such a chip even if it is isolated. An array of these chips, closely packed with a center-to-center spacing of one-half inch in order to reduce propagation delays, presents an even more difficult problem. The overall thermal resistance of the package for each individual chip must be less than 6”C/W to keep the temperature rise over ambient to less than 60°C. It is generally assumed that liquid cooling is required to handle such high power densities [ 11; thefefore, liquid cooling is generally being utilized in spite of an inherent aversion to using a liquid within an electrical system because of the potential damage to the circuits if a leak develops. For example, the IBM thermal conduction module, which is water cooled, has a thermal resistance of 8.4”C/W [2]. This paper describes a new compact air-cooled heat sink with the required thermal resistance. These sinks have been designed, built, and tested, and it has been demonstrated that they can handle 10 W chips on l/2 in centers. T THEORY The performance of a heat sink is usually characterized by its thermal resistance, 0 = AT/P, i.e., the rise in temperature Manuscript received August 15, 1983; revised November 21, 1983. The author is with the Sperry Corporation, P.O. BOX 500, MS ES-134, Blue Bell, PA 19424. where t is the thickness of the heat path, k is the thermal conductivity of the heat path material, and A is the crosssectional area of the heat path. &-“d can be made small by locating the sink close i:o the heat source and using material with large thermal conductivity. The average value of oheatis 1 oheat = - 2PCf where p is the density of the fluid, c is its specific heat, andf is the flow rate. &eatcan be reduced by using large flow rates; 8CO””will usually be the dominant component in a higher performance sink: eCO””= g 1 where h is the convective heat transfer coefficient and 11 is the area of the sink in contact with the cooling fluid. It has been generally recognized ihat this component can be reduced by increasing the surface area, typically with fins. The importance of increasing h, the convective- heat transfer coefficient has received less attention.. The heat transfer coefficient can be calculated from kf Nu h=---D where kf is the thermal conductivity of the fluid and D, the equivalent diameter of the channel, is given by D= 4 x flow area wetted perimeter . 0148-641 l/84/0300-0154$01 .OO 0 IEEE 1984 BLDBERG: NARROW CHANNEL Fig. 1. FORCED AIR HEAT SINK Heat sinking mounting. lr high aspect ratio channels D approximately equals twice : channel width w. Nu, the Nusselt number, is a dimensions parameter which is essentially defined by the above uation. Tuckerman and Pease [3] pointed out that the only way to :nificantly increase h is by decreasing D. They built a waterDied integral heat sink in a silicon integrated circuit and monstrated that extremely high power densities circuits uld be cooled in ‘that way. Our design for an air-cooled device based on the same principle. The value of Nu can be calculated exactly for certain ometries with laminar flow. The flow will be laminar if the :ynolds number is less than 2100. If the flow is laminar and ly developed, i.e., invariant along the direction of flow, it n be shown [41 that Nu = NIL, which is independent of the w velocity and depends only on the geometry of the channel. the entrance region where the temperature and velocity vary th in the direction of the flow and in the direction perpendicuto the flow, the Nusselt number is larger. [f we assume that Nu has its minimum asymptotic value, the imates will be conservative; h can be approximated by kf Nu, h=----- 2w Fig. 2. pointed out in [3], this result is consistent with an intuitive ldel in which the heat is conducted from the wall to the idle of the channel and is then carried away by the fluid w. The smaller the value of w, the greater is the thermal :dient and, therefore, the greater the heat removal. However ‘row channels increase the resistance to flow. The lower lit for w is set by the availability of blowers with the required ssure and flow. w(1 + o!) w2(1+ CY) 8CO””= ___ = 2zhS2 kfNu,S2z ’ To account for the finite conductivity of the sink, the area is reduced by a correction factor known as the fin efficiency [5]. The efficiency is given by tanh N 77= N ’ If S is much larger than the wall width, N can be approximated by [31 (+&I” where k, is the thermal conductivity of the wall. The value of z is used to set N and thereby 7. For N = 1 the heat removal by the tins is 76 percent of infinite length fins. For N=2, the heat removal is 96 percent of infinite length fins. Typically N is set to a value between 1 and 2. For N= N,,, z=N, k,,nw2 ~2 ~ kf Nu, * ( > Including the effect of finite conductivity, then DESIGN ?ig. 2 is a diagram of a narrow channel forced air heat sink. : square surface which is in contact with the substrate has esof length S. The channels are of width w and height z, and wall width is QW. Theaddition of the narrow channels to the sink multinlv --r -, the ---- channel forced air heat sink. area S2 by a factor of 2z/ w( 1 + a) if the heat transfer at the top and bottom of the channel is neglected. If one assumes infinite conductivity for the sink material and fully developed laminar flow, then N=z (c)“2zz * Narrow 0 W(Q l/2 + a _ l/2) conv = S2y&(kwkf Nu,) 1’2 where v. is the fin efficiency for N= NO. For a fixed w, econvis minimized by (Y = 1; i.e., wall width equal to channel width. The mean velocity of flow in high aspect ratio channels is u 156 IEEE TRANSACTIONS ON COMPONENTS. HYBRIDS, AND MANUFACTURING TECHNOLOGY, TABLE I VALUES FOR HEAT SINKS = w 2P/ 12~s where P is the pressure and p is the viscosity of the fluid [3]. The flow rate is j-z---= US.2 1 +a W (mils) I/2 w 3PNo ~ 12/.&Y”2+cY-“2) . Solving for the pressure we have (a 1 Pressure (” of H20) Reynolds Number 5 4.68 179 10 1.17 358 25 0.19 984 Bconv (“C/w) - 0.67 1.60 7.43 l/2+,-1/2), f = 30 l/mitt, Since the pressure to achieve the required flow at a fixed w is also minimized by CY= 1, wall width equal to the channel width, is the optimum choice. As a starting point in our design we set cheat=-= VOL. CHMT-7, NO. 1, MARC.H 1984 z = 0.5 in. CALCULATED TABLE II VALUES OF THERMAL RESISTANCE w(mils) 5 1“C/W. 2PCf l l Withthevalueforairofp = 1.13 x 10V3g/cm3andc = 1.01 W*s/g”C, f 1; 30 l/min = 1 CFM. The choice of w, which then fixed the pressure drop through the sink, was guided by the availability of commercial blowers. A reasonable value for the pressure is a few inches of water. In order to evaluate the narrow channel concept, sinks were made with S = 0.25 in (0.635 cm) and z = 0.5 in (1.27 cm). Three different values of w were used: 0.005 in (0.0127 cm), 0.010 in (0.0254 cm), and 0.025 in (0.0635 cm). Wall widths were equal to channel widths. The sinks were made of copper (k, = 3.91 W/cm”(J). The air fluid fixed k, = 2.85 x 10 -4 W/cm”C and p = 1.99 x 10 -4 g/cm*s. The sinks were fabricated from flat plate of the appropriate thickness cut to size. Shims of the same plate at the two ends of the sink fixed the channel width equal to the wall width. The ends were bonded by heli-arc welding. literslmin.) 0 (heat) = -!2ocf (f =30 t 0 (conductivity) = E Silicon Chip, t = 0.020” Alumina Substrate, t =0.030” Two Bonds Copper Sink Lead-in, t ~0.20” Total e (--C/W) 25 10 0.9 0.9 0.9 2.4 2.4 2.4 0.2 1.7 0.2 0.3 , l 0 (convection) 0.7 1.6 7.4 l f3 (total) 4.0 4.9 10.7 T I”‘3 A flow of 30 l/min with our geometry makes oheat = 0.9”U W. This flow rate corresponds to an average velocity of 1240 cm/s. The required pressure, the Reynolds number, Re = 2pv W/P, and Lnv were calculated for the three sinks and are displayed in Table I. A value of eight was used for Nu, for our high aspect ratio channels [6]. It should be noted that the Reynolds numbers are consistent with laminar flow (2100), and that the required pressures are moderate and readily available. 8 cond is the sum of the thermal resistance through 20 mils of silicon, 0.2”C/W, through the 30 mil alumina substrate, 1.7”C/W, through the copper lead-in of the sink, 0.3”C/W, and the two bonds which’ were estimated to be 0.1 “C/W each resulting in a total of 2.4 “C/W. These calculations are summarized in Table II. RESULTS The thermal resistance measurement set-up is shown in Fig. 1. A high-power chip eutectically bonded to a chip carrier was used as the heat source and temperature sensor. The forward voltage drop across a previously calibrated diode with fixed current is a reliable measure of the surface temperature of the chip. The heat sink was soldered to the other surface of the chip carrier. The temperature of the chip was measured for powers from 0 to 10 W with an air flow of 30 l/min, and the results for a 5, 10, and 25 mil channel width sink are shown in Fig. 3. AIR FLOW = 30 LiterslMin 0 2 4 6 8 IO POWER (Watts) Fig. 3. Temperature versus power for narrow channel heat sinks with air flow of 30 l/min. Measurements were also made at a constant power of 8 W while varying the air flow rate. Since oheatvaries inverse1.ywith the flow rate while econdand econvshould be independent of flow rate, a plot of temperature versus the reciprocal of the flow should be a straight line. The results are plotted for the 5 m.il and 10 mil sinks in Fig. 4. The linearity of the data confirms that the Nusselt number is constant and therefore the temperature profile is effectively fully developed. From these plots, one can extrapolate to infinite flow and determine 8 for oheat = 0. The results for all the measurements and calculations for 19 are collected in Table III. The measured value of 13are less than the calculated value in each case. This result is not unexpected because the calculations assume that the heat is removed only by the heat sink. In our single chip set-up a significant amount of heat is removed by natural convection directly and via the substrate to the ambient air. q #OLDBERG: NARROW CHANNEL 157 FORCED AIR HEAT SINK AIR FLOW (Liters per Minute) 30 20 14 I I I e (T/W) 70 Wc(mils) 65 lim f-m 5 2.0 10 2.5 “y/E----‘/ 0 .02 .04 l/AIR Temperature Fig. 4. COMPARISON B(‘CIW) Calculated (Liters versus reciprocal OF CALCULATED w (mils) FLOW POWER = 8 WATTS .06 .06 .lO per Minute)-’ of air flow at constant power of 8 W. TABLE III AND MEASUREDVALUES RESISTANCE (f =30Umin) Measured @C/W) Calculated OFTHERMAL (lim f - 00 ) Measured 5 4.0 3.4 3.1 2.0 10 4.9 3.7 25 10.7 5.5 4.0 - 2.5 - As a more realistic test of the heat sink, 16 chips were ounted on l/2 in centers on a 2 in x 2 in alumina substrate in 4 ~4 array. A heat sink with 10 mil channels was mounted rectly under each chip. An air plenum which supplied air >rna single blower to each of the sinks fits over the sinks. Fig. rhows two views of the plenum in position. A fine mesh screen covered the entrance to each sink. The ded flow resistance of the screen tended to equalize the air VWthrough each sink. Measurements confirmed that the air IWthrough each sink was equal to within 10 percent. Fig. 6 is )hotogiaph of the sink. side of the substrate before the top of : plenum was attached. Fig. 7 is an overall photograph of the ip side of the package. To insure low resistance bonds, the ips were eutectically bonded and the sinks were soldered to : substrate. Two test substrates were fabricated. One had only a single el of metallization so that the power dissipated in each chip lid not be individually controlled. Also it was necessary to asure the temperature of the chips with a less precise rared thermal gun because it was not possible to separately Lessthe diodes on the chip. ThS! air flow through each sink s 30 Umin. Our main interest was in the four chips in the lter of the substrate. The temperature, the power, and the rmal resistance for each of the four center sites are shown in . 8. The value of the thermal resistance was generally larger n for the individually mounted chips as expected but in good eement with the calculated value of 4.9 ‘C/W. ‘he other test substrate had multilayer metallization which mitted individual control of the power to each chip as well as of the more precise diode for temperature measurements. AIR ENTRANCE PLENUM INK C SUB&RATE (a) AIR PLENUM AIR SUPPLY PIPE \ Fig. 5. (aj Cross-sectional (b) view of air plenum mounted on substrate. (b) Plan view of air plenum mounted on substrate with air supply pipe. The multilayers added 28 pm of glass dielectric and some gold to the thermal path. The gold added negligible thermal resistante, but the glass increased Bcond by about 1.5 “C/W. Measurements were made with 10 W dissipated by each chip and an air flow of 30 Urnin through each sink. The temperature and the 158 Fig. 6. Fig. 7. Photograph Photograph of sink side of test packige. of chip side of test package with cabling. OLDBERG: NARROW CHANNEL 159 FORCED AIR HEAT SINK WIRE OR TAPE L---T/+ Fig. 10. :. 8. Temperature, power, and thermal resistance of four center chips with air flow of 30 Urnin each on substrate with single level of metalization. I- 2” - Configuration SINK for low Bcond. have a cut-out through which the sink with its pre-mounted and pretested chip would be inserted. A flange attached to the sink would provide a stop for inserting the sink as well as a surface to bond the sink to the substrate. CONCLUSION 85.7 “C 8.8“CIW 0 CHIP - \ 86.0 “C 5.9 “C/W ALUMINA SUBSTRATE In summary, a new style compact forced air heat sink has been designed, fabricated, and tested. It has been shown that narrow channel heat sinks using air as the cooling fluid can keep the temperature rise of 10 W chips on l/2 in centers below 60°C (6”C/W). Furthermore, by eliminating the substrate from the thermal path, thermal resistance can be reduced to about 3 “C/W. 83.4”C 5.6”CIW :. 9. Temperature and thermal resistance for four center chips with air flow ‘30 l/min and 10 W power each on substrate with multilayer metalization. :rmal resistance for the four center chips are shown in Fig. 9. le values are larger than those made on the substrate with a oglelevel of metallization because of the extra glass dielectric m, but the results are in good agreement with the calculated lues of 6.4”ClW. A significant improvement in the heat sink can be realized by :reasing econd.This could be done by removing the substrate Irn the thermal path. If the chip was bonded directly to the k, ticondwould be reduced to about 0.6”C/W, 0.2”C for the .con, 0.1 “C for the bond and 0.3 “C/W for the copper lead-in %e channels, reducing the calculated value of 19to 3.1 “C/W. ;. 10 shows how that could be done. The substrate would ACKNOWLEDGMENT I wish to thank Richard Melcher for the careful assembly of the test substrates and for assistance in collecting the data, and Carl Langheim for fabricating thk narrow channel heat sinks. I am also indebted to Joseph Tomei for his continuous support. REFERENCES [l] T. S. Steele, “Terminal and cooling requirements for LSI packages,” IEEE Trans. Components, Hybrids, Manufact. Technol., vol. [2] [3] [4] [5] [6] CHMT4, no. 2, pp. 187-191, June 1981. S. Oktay and H. C. Kammerer, “A conduction-cooled module for high performance LSI devices,” IBM J. Res. Develop., vol. 26, no. 1, pp. 55-60, Jan. 1982. D. B. Tuckerman and R. F. W. Pease, “High performance heat sinking for VLSI,” IEEEEIectron Dev. Letters, vol. EDL-2, no. 5, pp. 126129, May 1981. W. M. Kays, Convective Heat and Mass Transfer. New York: McGraw-Hill, 1958. M. Necate Ozisik, Basic Heat Transfer. New York: McGraw-Hill, 1977. J. G. Knudson and D. L. Katz, Fluid Dynamics and Heat Transfer. New York: McGraw-Hill, 1958.