FLASH INFORMATIQUE FI



expert On-chip cooling using micro-evaporators


La consommation d’énergie devenant un problème important pour les fabricants de matériels informatiques et les opérateurs de serveurs, de nouvelles solutions sont nécessaires pour réduire cette demande. L’amélioration des techniques de refroidissement des microprocesseurs est un des domaines où les plus grandes avancées sont possibles. L’expérience du LTCM dans le refroidissement par système bi-phasique des systèmes microélectroniques, longue de dix ans, est décrite. Les évaporateurs à multi-canaux fabriqués en silicium, cuivre ou aluminium peuvent refroidir de très hauts flux de chaleurs (jusqu’à 350 W/cm2), assurer la stabilité et l’uniformité des températures et permettre la gestion des points chauds sur un microprocesseur tout en étant compacts et en minimisant la quantité de liquide requise. Ce système de refroidissement peut réduire la consommation d’énergie de plus de 50%, avec d’autres gains possibles lorsque la chaleur est récupérée et utilisée à d’autres fins.



With energy consumption becoming a huge concern amongst hardware manufacturers and data center operators, novel solutions to reduce this consumption needs to come to light. Cooling of microprocessors is one of the largest areas where improvements can be gained. The LTCM lab’s decade of experience with two-phase cooling research for computer chips and power electronics will be described. Flow boiling in multi-microchannel cooling elements made of silicon, copper or aluminium have the potential to provide high cooling rates (up to as high as 350 W/cm2), stable and uniform temperatures, microprocessor hot spot management while being compact and thus minimizing the fluid inventory. This cooling technology can bring about savings of more than 50%, with further potential savings being made if the heat were to be recovered.


Jackson Braz MARCINICHEN

Jonathan OLIVIER

John Richard THOME


Cooling of data centers can represent up to 45% [1] of the total energy consumption using current cooling technologies (air cooling). In the US, this relates to an estimated 45 billion kWh usage by 2011 with an annual cost of $3.3 billion, or $648 billion with the inclusion of a carbon tax. And this is just for cooling. Most data centers make use of air-cooling technologies to ensure the correct running of the servers contained within. The limits of air-cooling, however, are being approached due to the performance increase in the microprocessors (CPU) in the servers, which will have heat fluxes in the order of 100 W/cm2 in the not too distant future. It was shown that air has a maximum heat removal capacity of about 37 W/cm2 [2]. The problem is made worse with servers being more densely packed, such as blade centers, racks that will be generating in excess of 60 kW of heat, while today’s data centers are designed for cooling capacities in the order of 10-15 kW per rack [3]. Hence, if data centers want to become green, other solutions to air-cooling are required.
One solution is to go to direct on-chip cooling. Recent publications show the development of primarily four competing technologies for on-chip cooling : microchannel single-phase flow, porous media flow, jet impingement and microchannel two-phase flow [4]. Leonard and Philips [5] showed that the use of such new technology for cooling of chips could produce savings in energy consumption of over 60%. Figure 1 shows the heat sink thermal resistances for diverse cooling technologies as a function of the pumping power to the dissipated thermal power ratio. The best heat sink solution should be that nearest the lower left corner because it represents the lowest thermal resistance at the lowest pumping power. It is clear that two-phase microchannel cooling is the best performing technology.

PNG - 5.7 ko
fig. 1
thermal resistance of heat sinks for diverse cooling technologies as a function of the pump to the dissipated power ratio.

The LTCM lab in particular has a very extensive research program on all important aspects of two-phase flow and two-phase heat transfer, where the primary eventual applications are the cooling of CPUs, power electronics, high energy physics particle detectors (CERN), micro-reactors, high power laser diodes, etc. This research is funded by Swiss FN, Nano-Tera, ESA, Eu, etc. projects as well as CTI, mandates and collaborations (IBM, ABB, Honeywell, etc.). Below, an overview of this work is given, emphasizing the research relevant to electronics cooling.
Figure 2 shows some typical microchannel coolers having channel widths in the range of 50 μm to 200 μm and fin heights ranging from 50 μm to 2 mm. The copper fins were manufactured by a process called micro deformation technology (MDT), a patented process of Wolverine Tube Inc. [6], a long-time partner of the lab.

PNG - 26.6 ko
fig. 2
examples of typical microchannel coolers in copper and silicon.

Similar fins can also be produced in aluminium, which is another common heat sink material. Such fins have been integrated into high performance micro-evaporators in collaboration with the LTCM lab. Also included is a silicon microchannel cooler having fin heights of 560 μm, fin width of 42 μm and a channel width of 85 μm. These channels were made by a process called deep reactive ion etching (DRIE) within a joint project with IBM.
It is worth noting that what happens in small channels in two-phase flows (evaporating refrigerant) is quite different than that for single-phase flows (water) in small channels. While initial studies in the literature reported significant size effects on friction factors and heat transfer coefficients in very small channels in single-phase flows, more accurate recent tests and analysis done with very smooth internal channels have shown that macroscale methods for single-phase flows work well at least down to diameters of 5-10 μm. This is not the case for macroscale two-phase flow methods, which usually do not work very well when compared to data for channels below about 2.0 mm diameter. Thus, it is inappropriate to extrapolate macroscale research results and methods to the microscale. Furthermore, many of the controlling phenomena and heat transfer mechanisms change when passing from macroscale two-phase flow and heat transfer to the microscale. For example, surface tension (capillary) forces become much stronger as the channel size diminishes while gravitational forces are weakened. Therefore, it is not physically sensible to refit macroscale methods to microscale data since the underlying physics has substantially changed, and thus dedicated research is required to investigate these microscale two-phase flows and develop new models and methods to describe them.
As a first view, Figure 3 depicts the buoyancy effect on elongated bubbles flowing in 2.0, 0.790 and 0.509 mm horizontal channels. In the 2.0 mm channel, the difference in liquid film thickness at the top compared to that at the bottom of the bubble is still quite noticeable (where this film thickness is the main resistance to heat transfer and thus cooling). Similarly, the film thickness in the 0.790 mm channel is still not uniform above and below the bubble. Instead, in the 0.509 mm channel, the film is now quite uniform. Interpreting these images and many others available in the lab, one ascertains that in small, horizontal channels that stratified types of flows (all vapor at top with liquid at the bottom) disappear. This transition is thus perhaps an indication of the lower boundary of macroscale two-phase flow, in this case occurring for a diameter somewhat greater than 2.0 mm. The upper boundary of microscale two-phase flow may be interpreted as the point in which the effect of gravity on the liquid-vapor interface shape becomes insignificant, such that the bubble in the 0.509 mm channel is thus a microscale flow, with the transition occurring at about this diameter at the present test conditions.

PNG - 5 ko
fig. 3
video images of slug (elongated bubble) flow in a 2.0, 0.8 and 0.5 mm horizontal channels with R-134a.

Heat transfer and pressure drop mechanisms are strongly affected by the type of flow patterns present in the channels, which need to be determined to develop prediction methods. With the aid of high-speed videography (videos up to 120’000 digital images per second) and laser photo diode signals, the LTCM was able to determine these regimes. Figure 4 shows a typical flow pattern map for R236fa inside a channel having a diameter of 1.03 mm. The flow patterns are identified as isolated bubbles (elongated bubbles), coalescing bubbles and annular flow (thin liquid film with vapor core).

PNG - 14.8 ko

Jacobi and Thome [8] proposed the first theoretically-based, elongated bubble (slug) flow boiling model for microchannels, modelling the thin film evaporation of the liquid film trapped between these bubbles and the channel wall and also accounting for the liquid-phase convection in the liquid slugs between the bubbles. The focus of their study was to demonstrate that the thin film evaporation mechanism was the principal heat transfer mechanism controlling heat transfer in slug flows in microchannels, not nucleate boiling as assumed by others in extrapolation of macroscale ideology to the microscale.
Figure 5 shows a representation of the three-zone model [9] of Thome et al. where Lp is the total length of the pair or triplet, LL is the length of the liquid slug, LG is the length of the bubble including the length of the dry wall of the vapor slug Ldry, and Lfilmis the length of the liquid film trapped by the bubble. The internal radius and the diameter of the tube are R and diwhile δo and δmin are the thicknesses of the liquid film trapped between the elongated bubble and the channel wall at its formation and at dry out of the film (only when dry out occurs). The evolution of successive bubbles is shown in the lower diagram. The local vapor quality, heat flux, microchannel internal diameter, mass flow rate and fluid physical properties at the local saturation pressure are input parameters to the model to predict the above parameters as well as the frequency of the bubbles, transient onset of local dryout, etc. The three-zone model predicts the local transient and time-averaged heat transfer rate at a fixed location along a microchannel during evaporation of a succession of elongated bubbles (with frequencies of passage as high as 900 Hz). The elongated bubbles are assumed to nucleate and quickly grow to the channel size upstream such that successive elongated bubbles are formed that are confined by the channel and grow in axial length, trapping a thin film of liquid between the bubble and the inner tube wall as they flow along the channel. The thickness of this film plays an important role in heat transfer. At a fixed location, the process is assumed to proceed as follows :

  1. a liquid slug passes,
  2. an elongated bubble passes (whose liquid film is formed from liquid removed from the liquid slug) and
  3. a vapor slug passes if the thin evaporating film of the bubble dries out before the arrival of the next liquid slug. The cycle then repeats itself upon arrival of the next liquid slug at a frequency f (=1/τ). Thus, a liquid slug and an elongated bubble pair or a liquid slug, an elongated bubble and a vapor slug triplet pass this fixed point at a frequency f that is a function of the formation and coalescence rate of the bubbles upstream.
PNG - 10.5 ko
fig. 5
three-zone heat transfer model of Thome et al. [9] for elongated bubble flow regime in microchannels. Top : Diagram illustrating a triplet comprised of a liquid slug, an elongated bubble and a vapor slug ; bottom : Bubble tracking of a triplet with passage of a new bubble at time intervals of τ.

For high heat flux cooling applications using multi-microchannel cooling channels, the critical heat flux (CHF) in saturated flow boiling conditions is a very important operational limit. It signifies the maximum heat flux that can be dissipated at the particular operating conditions by the evaporating fluid. Surpassing CHF means that the heated wall becomes completely and irrevocably dry, and is associated with a very rapid and sharp increase in the wall temperature due to the replacement of liquid by vapor adjacent to the heat transfer surface.
Revellin and Thome [10] have proposed the first theoretically based model for predicting critical heat flux in microchannels. Their model is based on the premise that CHF is reached when local dry out occurs during evaporation in annular flow at the location where the height of the interfacial waves matches that of the annular film’s mean thickness. To implement the model, they first solve one-dimensionally the conservation of mass, momentum and energy equations assuming annular flow to determine variation of the annular liquid film thickness δ along the channel. Then, based on the slip ratio given by the velocities of the two phases (liquid and vapor) and a Kelvin-Helmoltz critical wavelength criterion (assuming the height of the waves scales proportional to the critical wavelength), the wave height was modelled with the following empirical expression :

.

Then, when δ equals Δδ at the outlet of the microchannel, CHF is reached. Refer to Figure 6 for a simulation. The leading constant and two exponents were determined with a database including three fluids (R-134a, R-245fa and R-113) and three circular channel diameters (0.509 mm, 0.790 mm and 3.15 mm) taken from the LTCM CHF data of Wojtan et al. [11] and data from the Argonne Laboratory by Lazarek and Black [12]. Their model also satisfactorily predicted the Purdue R-113 data of Bowers and Mudawar [13] for circular multi-microchannels with diameters of 0.510 and 2.54 mm of 10 mm length. Furthermore, taking the channel width as the characteristic dimension to use as the diameter in their 1-d model, they were also able to predict the Purdue rectangular multi-microchannel data of Qu and Mudawar [14] for water. All together, 90% of the database was predicted within ±20%. As noted above, this model also accurately predicted the LTCM R-236fa multi-microchannel data of Agostini et al. [15]. Furthermore, this model also predicts CHF data of CO2 in microchannels from three additional independent studies as well as other fluids.

PNG - 5.6 ko
fig. 6
Revellin and Thome [10] CHF model showing the annular film thickness variation along the channel plotted versus the wave height.

Two-phase flow is much more complex than single-phase liquid flow and in this respect multi-microchannel flow boiling test sections can suffer from flow maldistribution, two-phase flow instabilities and even backflow effects. The flow may in fact flow back into the inlet header and some channels may become prematurely dry from too low of an inlet liquid flow rate.

PNG - 18.3 ko
fig. 7
dramatic effect that maldistribution can have on the heat transfer process [16].

Figure 7 shows a sequence of video images to demonstrate back flow and parallel channel instability in multi-microchannel test section (something to be avoided). A slug bubble was observed at the inlet of the topmost channel in (a). If the flow in the channel is pushed upstream by bubble growth downstream, the bubble goes back into the inlet plenum in (b), as there is no restriction at the channel inlet of the channel to prevent this. This reversed bubble quickly moves to one of the adjacent channels, (c), and breaks down into smaller parts before entering these channels, (d). Depending on its location, the inserted bubble becomes stagnant, (e) and (f), before moving forwards or backwards again.

PNG - 9.5 ko
fig. 8
flow boiling in a copper multi-microchannel test section from Park [16]showing the difference in bubble distribution a) without an inlet orifice and b) with an inlet orifice.

Using micro-inlet orifices can completely prevent backflow, flow instabilities and maldistribution. Figure 8a shows the maldistribution effect when no inlet orifice is used, with a large dry zone being visible in the top right corner. A critical heat flux of only 115W/cm2 was achieved. Figure 8b shows that maldistribution is avoided when making use of an inlet orifice, with heat fluxes in excess of 350W/cm2 being obtainable. This is equivalent to cooling of 35’000 100-Watt light bulbs per meter squared of surface area ! Hence, microscale flow boiling can dissipate very high heat fluxes !
The LTCM has thus proven that heat transfer characteristics of microchannels can be predicted successfully while also showing the capability microchannels have in removing high heat fluxes. Numerous experimental campaigns on multi-microchannel evaporators have been performed by the LTCM.
It was proven that, using a refrigerant evaporating at 60°C (such as refrigerant R134a that is circulating in your climatisation system of your automobile), microprocessors could be kept well below their 85°C limit, while removing heat fluxes in excess of 180W/cm2 or more [17]. Two-phase flow is also ideally suited for cooling of electronic hotspots (local heat fluxes on CPUs up to 400 W/cm2 or more) as heat transfer coefficients (thermal resistances) naturally increase (decrease) over hotspot locations in the flow boiling process described earlier. This has the implications that electronics can have a more uniform temperature, implying that problems associated with adverse temperatures gradients are greatly diminished and hence higher clock speeds can be utilized. The power requirements for removing the heat is also considerably lower than required for traditional air-cooling methods. This is due to the much larger heat carrying capacity refrigerants have over air, while also taking advantage of its latent heat, which is much higher than sensible heat of single-phase fluids. Savings of over 50% are achievable using this technology, with extra savings being possible if the energy gained from cooling the processors were to be used in a secondary process. This is shown in Figure 9.
Numerous other aspects of these microscale two-phase flows are under investigation in the LTCM lab, such as micro-PIV to characterize flow (with cavitation) through micro-orifices as small as 15-25 microns, transient aspects of the evaporation and bubble coalescence process, time-strip analysis of high speed videos to discern information on the dry out process and wave formation, flow pattern transition theory, etc.

PNG - 10.3 ko
fig. 9
potential energy savings for a data center when using on-chip cooling with energy recovery compared to traditional air cooling.

References

[1] Koomy, J.G., Estimating Regional Power Consumption by Servers : A Technical Note. 2007, Lawrence Berkeley National Laboratory : Oakland, CA.
[2] Saini, M. and Webb, R.L., Heat Rejection Limits of Air Cooled Plane Fin Heat Sinks for Computer Cooling. IEEE Transactions on Components and Packaging Technologies, 2003. 26(1) : p. 71-79.
[3] Samadiani, E., Joshi, S. and Mistree, F., The Thermal Design of a Next Generation Data Center : A Conceptual Exposition. Journal of Electronic Packing, 2008. 130 : p. 041104-1 - 041104-8.
[4] Agostini, B., Fabbri, M., Park, J.E., Wojtan, L., Thome, J.R. and Michel, B., State-of-the-art of High Heat Flux Cooling Technologies. Heat Transfer Engineering, 2007. 28 : p. 258-281.
[5] Leonard, P.L. and Phillips, A.L., The Thermal Bus Opportunity - A Quantum Leap in Data Center Cooling Potential in ASHRAE Transactions. 2005. Denver, CO.
[6] Wolverine Tube Inc. 2010.
[7] Ong, C.L., Macro-to-Microchannel Transition in Two-phase flow and Evaporation, in Heat and Mass Transfer Laboratory (LTCM). 2010, École Polytechnique Fédérale de Lausanne : Lausanne.
[8] Jacobi, A.M. and Thome, J.R., Heat transfer model for evaporation of elongated bubble flows in microchannels. Journal of Heat Transfer-Transactions of the Asme, 2002. 124(6) : p. 1131-1136.
[9] Thome, J.R., Dupont, V. and Jacobi, A.M., Heat transfer model for evaporation in microchannels. Part I : presentation of the model. International Journal of Heat and Mass Transfer, 2004. 47 : p. 3375-3385.
[10] Revellin, R. and Thome, J.R., An analytical model for the prediction of the critical heat flux in heated microchannels. Int. J. Heat Mass Transfer, 2008. 51 : p. 1216-1225.
[11] Wojtan, L., Revellin, R. and Thome, J.R., Investigation of saturated critical heat flux in a single uniformly heated microchannel. Experimental Thermal and Fluid Science, 2006. 30 : p. 765-774.
[12] Lazarek, G.M. and Black, S.H., Evaporating Heat Transfer, Pressure Drop and Critical Heat Flux in a Small Vertical Tube with R-113. International Journal of Heat and Mass Transfer, 1982. 25(7) : p. 945-960.
[13] Bowers, M.B. and Mudawar, I., High flux boiling in low flow rate, low pressure drop mini-channel and micro-channel heat sinks. Int. J. Heat Mass Transfer, 1994. 37 : p. 321-332.
[14] Qu, W. and Mudawar, I., Measurement and correlation of critical heat flux in two-phase micro-channel heat sink. Int. J. Heat Mass Transfer, 2004. 47 : p. 2045-2059.
[15]Agostini, B., Revellin, R., Thome, J.R., Fabbri, M., Michel, B., Calmi, D. and Kloter, U., High heat flux flow boiling in silicon multi-microchannels - Part III : Saturated critical heat flux of R236fa and two-phase pressure drops. International Journal of Heat and Mass Transfer, 2008. 51(21-22) : p. 5426-5442.
[16] Park, J.E., Critical Heat Flux in Multi-Microchannel Copper Elements with Low Pressure Refrigerants. 2008, École Polytechnique Fédérale de Lausanne.
[17] Madhour, Y., Olivier, J.A., Costa-Patry, E., Paredes, S., Michel, B. and Thome, J.R., Flow Boiling of R134a in a Multi-Microchannel Heat Sink with Hotspot Heaters for Energy-Efficient Microelectronic CPU Cooling Applications. IEEE Transactions on Components and Packaging Technologies, Accepted for publication, 2010.



Cherchez ...

- dans tous les Flash informatique
(entre 1986 et 2001: seulement sur les titres et auteurs)
- par mot-clé

Avertissement

Cette page est un article d'une publication de l'EPFL.
Le contenu et certains liens ne sont peut-être plus d'actualité.

Responsabilité

Les articles n'engagent que leurs auteurs, sauf ceux qui concernent de façon évidente des prestations officielles (sous la responsabilité du DIT ou d'autres entités). Toute reproduction, même partielle, n'est autorisée qu'avec l'accord de la rédaction et des auteurs.


Archives sur clé USB

Le Flash informatique ne paraîtra plus. Le dernier numéro est daté de décembre 2013.

Taguage des articles

Depuis 2010, pour aider le lecteur, les articles sont taggués:
  •   tout public
    que vous soyiez utilisateur occasionnel du PC familial, ou bien simplement propriétaire d'un iPhone, lisez l'article marqué tout public, vous y apprendrez plein de choses qui vous permettront de mieux appréhender ces technologies qui envahissent votre quotidien
  •   public averti
    l'article parle de concepts techniques, mais à la portée de toute personne intéressée par les dessous des nouvelles technologies
  •   expert
    le sujet abordé n'intéresse que peu de lecteurs, mais ceux-là seront ravis d'approfondir un thème, d'en savoir plus sur un nouveau langage.