Skip to content
LumenaryAcademy
Frontier 85 min

Thermal management for computing

Why chips get hot, and how a kilowatt leaves a few square centimetres of silicon — the thermal resistance chain, heat sinks, heat pipes and vapour chambers, cold plates and immersion, two-phase cooling, and the water, energy and reusable heat of AI data centres.

Code tested · 6AI-draftedAwaiting review by a mechanical engineer

Your guides: Chloe · The Designer, Margaret · The Test Engineer, Gideon · The Checker

MotivationH

A kilowatt on a postage stamp

A kitchen hob ring turns 2 kW into heat over about 200 cm², some 10 W/cm², and glows red doing it. An AI accelerator of 2026 turns about 1 kW into heat in a package whose silicon has an area of perhaps 16 cm², over 60 W/cm² on average and several times that at local hot spots, and it must keep that silicon below about 90–105 °C, ideally much cooler, for years. As Hardware for AI in Lumenary EE describes, one rack of 72 such accelerators draws about 120 kW, ten times a typical rack of a decade earlier, and designs of several hundred kilowatts per rack are announced for 2027.

Every one of those watts becomes heat. Getting it out is a mechanical engineering problem from end to end: conduction through layers a fraction of a millimetre thick, convection into air or water, phase change in heat pipes and boiling liquids, pumps, pipes and heat exchangers, and finally rejection to the outdoors or to a city's heating network. The Lumenary EE lesson on powering data centres follows the electricity in and introduces PUE and WUE. This lesson follows the heat out, with the tools of heat transfer and viscous flow.

IntuitionH

Why chips get hot, and why it matters

A CMOS chip's power is mostly dynamic power, P≈αCV2fP \approx \alpha C V^2 f: every clock cycle a fraction α\alpha of its switched capacitance CC is charged to the supply voltage VV and discharged again, ff times a second; to that is added leakage through transistors that are nominally off. For decades, shrinking transistors lowered CC and VV together and kept power per unit area constant. Since about 2005 the supply voltage has barely fallen, so packing in more transistors raises the power density; transistor scaling tells that story. AI accelerators add a second push: they keep thousands of multipliers busy at once, and their makers gain performance by spending more power.

Temperature matters for three reasons:

  • Reliability. Wear-out mechanisms in chips, such as electromigration in the wiring, speed up roughly exponentially with temperature (Arrhenius behaviour), so a few kelvin can change lifetime noticeably.
  • Leakage. Leakage current grows with temperature, which raises power, which raises temperature: a positive feedback that cooling must keep stable.
  • Performance. Modern processors measure their own temperature and slow their clocks when it climbs. Better cooling buys sustained speed.
DefinitionsH

The thermal resistance chain

Definition— Thermal resistance

For a steady heat flow QQ through a path whose ends differ in temperature by ΔT\Delta T, the thermal resistance is R=ΔT/QR = \Delta T/Q, in K/W. A plane layer of thickness tt, conductivity kk and area AA has R=t/(kA)R = t/(kA); a surface with convection coefficient hh has R=1/(hA)R = 1/(hA); a contact between two surfaces has an interface resistance R′′/AR''/A, where R′′R'' (in m²·K/W, often quoted in cm²·K/W) characterises the joint. Resistances in series add.

The heat from a transistor's junction crosses a chain of layers, each with its own resistance. The package makers' standard metrics, defined by JEDEC's JESD51 series, include θJC\theta_{JC} (junction to case) and θJA\theta_{JA} (junction to ambient, in a specified test set-up).

substratesilicon dieTIM1copper lidTIM2cold plate with fin channelscoolant inout, warmerjunction, T_jcoolant inlet, T_insiliconTIM1lidTIM2plate baseconvectioncoolant warmingQ
A lidded chip package on a liquid-cooled cold plate, and its thermal resistance network. Heat from the junction crosses the silicon, a thermal interface material (TIM1), the copper lid, a second interface (TIM2) and the cold plate's base, then passes by convection into the coolant, which warms as it flows. Removing the lid removes two layers of the chain.

The program below assembles the chain for a 1 kW accelerator with a lid and a cold plate. Every value is illustrative but in the range published for such packages: a good polymer interface material of 0.10 cm²·K/W, a finned copper cold plate with an effective heat-transfer coefficient of 10510^5 W/(m²·K) over its footprint, and coolant (25% propylene glycol in water) flowing at 1.5 L/min per kilowatt. Hot spots are modelled crudely: where the local heat flux is three times the average, the two layers nearest the die see three times the temperature gradient.

python
# Junction-to-coolant thermal resistance chain of a liquid-cooled accelerator (illustrative values)
A_die, A_plate = 16e-4, 25e-4                # m^2: silicon area; area the lid and cold plate spread over
R_tim = 0.10e-4                              # m^2 K/W (0.10 cm^2 K/W): a good polymer thermal interface
lidded = [  # (layer, resistance in K/W)
    ("silicon die, 0.5 mm",          0.5e-3 / (130 * A_die)),
    ("TIM1, die to lid",             R_tim / A_die),
    ("copper lid, 2 mm",             2e-3 / (390 * A_die)),     # 1-D through the die footprint: pessimistic
    ("TIM2, lid to cold plate",      R_tim / A_plate),
    ("cold plate base, 2 mm Cu",     2e-3 / (390 * A_plate)),
    ("convection to the coolant",    1 / (1.0e5 * A_plate)),    # effective h of a finned plate, per footprint
]
R = sum(r for _, r in lidded)
Q, lpm_per_kW, rho, cp = 1000.0, 1.5, 1020.0, 3900.0           # coolant: 25 % propylene glycol in water
m_dot = Q / 1000 * lpm_per_kW / 60 / 1000 * rho
rise = Q / (m_dot * cp)                                         # coolant temperature rise across the plate
print(f"{'layer':30s} {'R (K/W)':>8s}  {'dT at 1 kW':>10s}")
for name, r in lidded:
    print(f"{name:30s} {r:8.4f}  {Q * r:8.1f} K")
print(f"{'conduction + convection':30s} {R:8.4f}  {Q * R:8.1f} K")
print(f"coolant at {lpm_per_kW} L/min per kW warms by {rise:.1f} K (mean junction sees half of it)")

# Hot spots: if the local heat flux is 3x the average, the layers nearest the die see 3x the gradient
hot = 3.0
def junction_rise(P, R_chain, R_near):
    return P * R_chain + (hot - 1) * P * R_near + rise / 2      # flow scaled with power: rise is constant

near = lidded[0][1] + lidded[1][1]
lidless = [lidded[0][1], R_tim / A_die, lidded[4][1], lidded[5][1]]   # cold plate on the bare die
print(f"\nlidless stack (die, one TIM, plate, convection): {sum(lidless):.4f} K/W\n")
print(" power (W)   junction rise (K)      highest coolant inlet for a 95 C junction")
print("             lidded   lidless       lidded    lidless")
for P in (700, 1000, 1400, 2000):
    a, b = junction_rise(P, R, near), junction_rise(P, sum(lidless), near)
    print(f"  {P:6d}    {a:6.1f}   {b:6.1f}        {95 - a:6.1f} C  {95 - b:6.1f} C")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

The chain totals 0.022 K/W, about 22 K at 1 kW, and the largest single term is the thin interface between die and lid: 0.1 mm of material costs more temperature than 2 mm of copper. That is why the industry fights over interface materials (indium and other metal solders, liquid metals, graphite and carbon-nanotube pads) and why many accelerators now ship lidless, with the cold plate pressed onto the bare die through one interface. Here that saves 0.007 K/W.

The bottom table is the heart of the matter. The junction must stay below its limit (95 °C here), so the chain fixes the warmest coolant the chip can accept. At 700 W the lidded package tolerates coolant at 62.5 °C; at 2 kW, only 11.5 °C, which means mechanical refrigeration all year in any climate, unless the stack improves. The lidless stack buys back about 14 K at 2 kW. Every kelvin of resistance removed from the chain becomes a kelvin of warmer coolant, and, as the last computation in this lesson shows, warmer coolant is what makes chiller-free cooling, low water use and heat reuse possible.

ComputationC

Air: heat sinks, fins and fans

Most of the world's chips, from phones to laptops to the CPUs in most servers, are still cooled by air. A heat sink spreads heat from a small base into many thin fins, multiplying the area exposed to the air; a fan pushes air between them. Two classic results from heat transfer decide its performance. A fin of height HH and thickness tft_f conducts heat outwards while losing it from its faces, so its tip is cooler than its root, and its fin efficiency is ηf=tanh⁡(mH)/(mH)\eta_f = \tanh(mH)/(mH) with m=2h/(ktf)m = \sqrt{2h/(k t_f)}. And the air warms as it absorbs heat: with heat-capacity rate C=ρQcpC = \rho Q c_p and hAhA the effective conductance of the fins, the heat sink's resistance from base to inlet air is R=1/[C (1−e−hA/C)]R = 1/\big[C\,(1 - e^{-hA/C})\big], which can never be less than 1/C1/C.

More fins add area and narrow the gaps, which raises the heat-transfer coefficient, but narrow gaps also resist the flow, and a fan delivers less air against more resistance. The flow settles where the fan curve meets the system curve, exactly as for the pumps in viscous flow and pipes. The program finds the best number of fins for a 100 mm × 100 mm aluminium heat sink with 40 mm fins, cooled by a fan whose curve is a simple parabola, carrying 300 W.

python
import numpy as np

# Plate-fin heat sink in a duct, blown by a fan: how many fins?
W, L, H, tf = 0.100, 0.100, 0.040, 0.8e-3      # base width, flow length, fin height, fin thickness (m)
k_al = 200.0                                   # aluminium alloy, W/(m K)
rho, cp, mu, k_air, Pr = 1.16, 1007.0, 1.85e-5, 0.0263, 0.71   # air near 30 C
dp_max, Q_max = 150.0, 0.015                   # fan curve: dp = dp_max (1 - (Q/Q_max)^2), Pa and m^3/s
P_chip = 300.0                                 # W

def heat_sink(N):
    s = (W - N * tf) / (N - 1)                 # gap between fins
    A_flow = (N - 1) * s * H
    Dh = 2 * s * H / (s + H)                   # hydraulic diameter of one channel
    def system_dp(Q):                          # laminar friction (parallel plates, f Re = 96) + entry and exit
        V = Q / A_flow
        return (96 * mu / (rho * V * Dh) * L / Dh + 1.5) * 0.5 * rho * V**2
    lo, hi = 1e-6, Q_max                       # operating point: fan curve meets system curve (bisection)
    for _ in range(60):
        Q = 0.5 * (lo + hi)
        lo, hi = (Q, hi) if dp_max * (1 - (Q / Q_max) ** 2) > system_dp(Q) else (lo, Q)
    V = Q / A_flow
    Re = rho * V * Dh / mu
    Gz = Dh / L * Re * Pr                      # Graetz number: how far from fully developed
    h = (7.54 + 0.03 * Gz / (1 + 0.016 * Gz ** (2 / 3))) * k_air / Dh   # laminar, entry region included
    m = np.sqrt(2 * h / (k_al * tf))
    eta_f = np.tanh(m * H) / (m * H)           # fin efficiency
    hA = h * (eta_f * 2 * N * H * L + (W - N * tf) * L)
    C = rho * Q * cp                           # heat-capacity rate of the air, W/K
    R = 1 / (C * (1 - np.exp(-hA / C)))        # base to inlet air, including the air's own warming
    return s, Q, Re, h, eta_f, system_dp(Q), P_chip / C, R

print(" fins  gap(mm)  flow(L/s)    Re   h(W/m2K)  fin eff  dp(Pa)  air rise(K)  R(K/W)")
results = {}
for N in (20, 25, 30, 35, 40, 45, 50, 55, 60, 70):
    s, Q, Re, h, eta, dp, rise, R = heat_sink(N)
    results[N] = R
    print(f" {N:4d}  {s * 1e3:6.2f}   {Q * 1e3:7.2f}  {Re:5.0f}  {h:7.1f}   {eta:6.3f}  {dp:5.0f}  {rise:9.1f}   {R:7.4f}")
N_best = min(results, key=results.get)
print(f"\nbest: {N_best} fins, {results[N_best]:.3f} K/W, so the base runs "
      f"{results[N_best] * P_chip:.0f} K above the inlet air at {P_chip:.0f} W")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

From 20 to 50 fins the resistance falls from 0.25 to 0.098 K/W: more area, and a higher hh in the narrower gaps. Beyond about 50 fins the fan can no longer push enough air through the 1 mm gaps, the flow collapses, and the resistance rises again. At the optimum the air leaving the fins is already 25 K warmer than it entered, close to the 1/C1/C limit: the heat sink is running out of air, not of area.

ComputationC

Heat pipes, vapour chambers and boiling

A heat pipe is a sealed tube, usually copper with water inside at low pressure, lined with a wick of sintered copper powder, grooves or mesh. At the hot end (the evaporator) the water boils; the vapour rushes to the cold end (the condenser), condenses and gives up its latent heat; and capillary forces in the wick pull the liquid back. Because it moves heat as latent heat with only a small temperature drop, a heat pipe can have an effective conductivity tens to hundreds of times that of solid copper. A vapour chamber is a flat heat pipe that spreads heat in two dimensions under a hot chip, and is common under high-power processors, in phones and in gaming laptops.

evaporatoradiabatic sectioncondenserheat inheat outvapourwick: liquid returns by capillary action
A heat pipe. Heat entering the evaporator boils the working fluid; vapour flows along the core to the condenser, where it condenses and releases its latent heat; the wick lining the wall pumps the liquid back by capillary action. The capillary pressure must overcome the friction of the liquid in the wick, the vapour's pressure drop and any adverse gravity.

Two limits matter most. The capillary limit: the wick's maximum capillary pressure, 2σ/reff2\sigma/r_\text{eff} for surface tension σ\sigma and effective pore radius reffr_\text{eff}, must drive the liquid through the wick (Darcy flow, with permeability KK) against any gravity head. The working fluid enters through its merit number ρlσhfg/μl\rho_l \sigma h_{fg}/\mu_l; water's is the highest of any common fluid between about 30 and 200 °C. And the boiling limit: a heated surface under a pool of liquid can only shed heat by boiling up to the critical heat flux (CHF), beyond which a vapour blanket insulates it and its temperature jumps. Novak Zuber's hydrodynamic theory of 1958 gives

qmax′′=C hfg ρv1/2[σg(ρl−ρv)]1/4,C≈0.131,q''_\text{max} = C\,h_{fg}\,\rho_v^{1/2}\big[\sigma g (\rho_l - \rho_v)\big]^{1/4}, \qquad C \approx 0.131,

for a large flat heater.

python
import math
g = 9.81

# 1. Capillary limit of a copper-water heat pipe with a sintered wick (illustrative geometry)
rho_l, mu_l, sigma, h_fg = 983.0, 4.67e-4, 0.0662, 2.358e6      # water at 60 C
merit = rho_l * sigma * h_fg / mu_l                                # the liquid's figure of merit, W/m^2
K, r_eff = 3e-11, 40e-6                                          # wick permeability (m^2), pore radius (m)
A_wick = math.pi / 4 * (5.4e-3**2 - 4.4e-3**2)                   # 0.5 mm wick inside a 6 mm pipe
L_eff, L_total = 0.15, 0.20                                      # effective and total length (m)
p_cap = 2 * sigma / r_eff
print(f"merit number of water at 60 C: {merit:.2e} W/m^2;  capillary pressure {p_cap:.0f} Pa")
print(" tilt (evaporator above condenser)   gravity head (Pa)   capillary limit (W)")
for deg in (0, 15, 30, 60, 90):
    p_grav = rho_l * g * L_total * math.sin(math.radians(deg))
    q_max = (p_cap - p_grav) * rho_l * h_fg * K * A_wick / (mu_l * L_eff)   # Darcy flow in the wick
    print(f"   {deg:3d} degrees                           {p_grav:6.0f}           {max(q_max, 0):6.1f}")

# 2. Critical heat flux in pool boiling (Zuber): q_max = C h_fg rho_v^0.5 [sigma g (rho_l - rho_v)]^0.25
fluids = {  # rho_l, rho_v (kg/m^3), h_fg (J/kg), sigma (N/m), at boiling under 1 atm, approximate
    "water, 100 C": (958.4, 0.598, 2.257e6, 0.0589),
    "fluorinated dielectric, ~56 C": (1600.0, 13.4, 9.48e4, 0.0083),
}
print("\n fluid                            CHF (W/cm^2)")
for name, (rl, rv, hfg, sig) in fluids.items():
    chf = 0.131 * hfg * math.sqrt(rv) * (sig * g * (rl - rv)) ** 0.25
    print(f" {name:32s} {chf / 1e4:8.1f}")
print(f" average flux of 1 kW from 16 cm^2 of silicon: {1000 / 16:.1f} W/cm^2")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

The 6 mm heat pipe carries about 25 W lying flat, but only 11 W standing upright with its hot end at the top, because the wick must then lift the liquid 20 cm against gravity. That is why a laptop's cooling is designed for a known orientation, and why phones use thin vapour chambers whose capillary paths are short.

The boiling numbers explain a constraint on two-phase immersion. Water can boil away about 110 W/cm² from a flat surface, but water conducts electricity, so it cannot touch the chips. The dielectric fluids that can, mostly fluorinated, have small latent heats and surface tensions, and their critical heat flux is only about 15 W/cm², a quarter of the accelerator's average flux. Two-phase immersion of high-power chips therefore relies on boiling enhancement: porous coatings, sintered or finned "boiler plates" attached to the chip that multiply the boiling area and the nucleation sites.

ExplanationH

Liquid cooling for AI data centres

Direct-to-chip cooling, the mainstream for AI racks as of 2026, puts a cold plate on each accelerator and CPU. Its fins are fine channels, typically a few tenths of a millimetre wide, machined, skived (sliced up from the copper base) or made by additive manufacturing, and the coolant is usually water with propylene glycol and corrosion inhibitors, carefully filtered so that the channels do not clog. The idea goes back to 1981, when David Tuckerman and Fabian Pease at Stanford etched microchannels into the back of a silicon chip and removed 790 W/cm². In the rack, manifolds and drip-free quick disconnects connect each server to a coolant distribution unit (CDU), which pumps the server loop and exchanges its heat with the building's facility water through a plate heat exchanger. Fans still cool memory, power supplies and network cards: typically most of a rack's heat, but not all, goes to liquid.

server, cold platesserver, cold platesserver, cold platesserver, cold platesAI rackCDUpump, heat exchangerwarm returncool supplyfacility waterdry coolerchiller or evaporative assiston hot daysheat reusedistrict heating, via a heat pump
The heat path of a liquid-cooled AI rack. Cold plates on the chips feed a technology cooling loop pumped by a coolant distribution unit (CDU). A heat exchanger in the CDU passes the heat to the facility water loop, which rejects it outdoors through dry coolers (with chillers or evaporative assistance on hot days) or delivers it to a heat-reuse customer.

The alternatives, as of 2026:

  • Single-phase immersion: whole servers submerged in tanks of a dielectric oil (hydrocarbon or synthetic ester) that is pumped through a heat exchanger. Simple and quiet, with no fans, but heavy, messy to service, and the fluids can affect cable and label materials. Used at some sites, notably for cryptocurrency mining and some edge and high-density installations.
  • Two-phase immersion: servers in a bath of a fluid that boils on the chips and condenses on coils above. Excellent heat transfer, but it has depended on fluorinated fluids. 3M, a major supplier, ended production of its fluorinated fluids at the end of 2025 as part of its exit from PFAS manufacturing, and several proposed replacements also fall within broad regulatory definitions of PFAS, which regulators in Europe and elsewhere are still settling (see sustainable chemistry in Lumenary Chemistry). The approach remains a niche.
  • Two-phase cold plates: a refrigerant boils inside the cold plate itself, pumped round a closed loop. Boiling absorbs heat at nearly constant temperature, and the flow needed is several times lower than for water, which is attractive for the hottest chips; the challenges are flow distribution between many parallel plates, instabilities, and the refrigerants' own environmental rules. Several start-ups and established cooling firms offer them, and they are a leading candidate for racks of several hundred kilowatts.
  • Cooling inside the chip: coolant channels etched into the silicon or into the package, as Tuckerman and Pease first did, and as recent research has developed further (van Erp and colleagues at EPFL, 2020). Microsoft reported a prototype in 2025 with channels etched into the back of a chip. It removes most of the chain at once, at the cost of new manufacturing and reliability questions.

Standards are catching up. ASHRAE's Technical Committee 9.9 publishes classes for liquid cooling named by the facility water supply temperature, and the Open Compute Project publishes open specifications for cold plates, CDUs, quick disconnects and immersion tanks, so that equipment from different makers can work together.

Note

The facility side of liquid cooling involves pumps, pressurised pipework, heat exchangers and refrigeration plant, as well as the high-voltage electrical systems described in Lumenary EE. It is designed, installed and maintained by qualified engineers and technicians under pressure-equipment, refrigerant-handling and electrical-safety rules. Experiments with cooling at home are best kept to a small fan and heat sink on a low-voltage project board.

ComputationC

Water, energy and reusable heat

The heat ends up outdoors, or in someone's radiator. How much energy and water that costs depends almost entirely on the temperature at which it is rejected. A dry cooler (a radiator with fans) can cool the facility water to within a few kelvin of the outdoor air, using only fan power. When the air is too warm, sites add evaporative cooling, which can cool towards the lower wet-bulb temperature but consumes water, or mechanical chillers, which use electricity (the refrigeration cycles of power and refrigeration cycles). The program counts the hours in a year when a dry cooler alone can supply coolant at a given temperature, for two synthetic climates.

python
import numpy as np

# A synthetic year of hourly air temperatures: annual mean, a seasonal swing and a daily swing (deg C)
hours = np.arange(8760)
day, hour = hours / 24, hours % 24
def climate(mean, seasonal, daily, hottest_day=200):
    return (mean + seasonal * np.cos(2 * np.pi * (day - hottest_day) / 365)
            + daily * np.cos(2 * np.pi * (hour - 15) / 24))

climates = {"cool temperate": climate(9, 8, 4), "hot": climate(27, 8, 6)}
approach = 6.0          # a dry cooler can bring the coolant to within about 6 K of the air temperature

print(" coolant supply   hours a year a dry cooler alone suffices")
print("   temperature    " + "   ".join(f"{name:>16s}" for name in climates))
for supply in (18, 25, 32, 40, 45):
    shares = [np.mean(T <= supply - approach) for T in climates.values()]
    print(f"     {supply:3d} C       " + "   ".join(f"{s:15.0%} " for s in shares))
for name, T in climates.items():
    print(f"{name}: coldest hour {T.min():.0f} C, hottest {T.max():.0f} C")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

In the cool climate, coolant at 25 °C can come from dry coolers 94% of the year and at 32 °C all year. In the hot climate, 18 °C coolant is never possible without chillers, while 40 °C coolant is available 81% of the hours and 45 °C 95%. Put that next to the first program: a chip stack that tolerates 45 °C coolant can be cooled almost anywhere without chillers or evaporation, and one that needs 11 °C cannot be cooled anywhere without them. Thermal resistance in the package becomes energy and water at the site.

Water. Evaporating water absorbs about 2.4 MJ/kg, so rejecting 1 kWh of heat purely by evaporation uses roughly 1.5 litres; the EE lesson computes what that means for a campus. Closed-loop liquid cooling with dry coolers uses almost no water on site, but more fan and, on hot days, chiller electricity, and the power stations supplying that electricity may themselves use water. The honest comparison is site-specific: the local climate, the stress on the local watershed, and the source of the electricity.

Heat reuse. Return water from cold plates at 40–60 °C is warm enough for low-temperature district heating directly, or for a heat pump to lift it to 70–90 °C much more efficiently than from outdoor air. Data centres already feed district heating in Odense (Denmark), Stockholm and the Helsinki region, among other places, and Germany's Energy Efficiency Act of 2023 requires new data centres to reuse a growing share of their heat. The limits are practical: a heat customer close by, heat demand that is seasonal while the data centre's output is constant, and contracts between parties with very different time horizons.

ExplanationH

The professional view

Thermal engineers for electronics work at every scale: package and interface materials with the chip makers; heat sinks, cold plates and vapour chambers with specialist suppliers in Taiwan, China, Japan, Europe and the United States; whole servers and racks with conjugate heat-transfer CFD tools from several vendors; and data-centre halls with airflow and facility-loop models. Design rules come from JEDEC for package metrics, ASHRAE TC 9.9 for environmental classes, and the Open Compute Project for open hardware specifications. Every design ends in testing: thermal test vehicles with programmable heaters, infrared imaging, flow and pressure-drop rigs, and long reliability runs with thermal cycling, because interface materials and pumps age.

The trends to watch, as of late 2026: rack powers climbing from about 120 kW towards several hundred kilowatts, with higher-voltage power delivery (described in Lumenary EE); warmer facility water, to cut chillers and enable reuse; two-phase cold plates for the hottest chips; cooling built into packages and chips; and growing regulatory attention to data centres' energy, water and heat reuse in the EU and elsewhere. Claims of dramatic gains deserve the usual questions: compared with what baseline, at what coolant temperature, and measured how?

Common misconceptionsH

Common misconceptions

Misconception

"More thermal paste conducts more heat." Interface materials conduct far worse than metal; their job is only to fill the microscopic gaps between two surfaces. The thinnest continuous layer is best, which is why bond-line thickness is controlled to tens of micrometres.

Misconception

"Colder coolant is always better." It lowers the junction temperature, but chilling water costs electricity and often water, rules out free cooling for much of the year and makes heat reuse impossible. The goal is coolant as warm as the chip allows, which is why reducing the package's thermal resistance is so valuable.

Misconception

"Liquid-cooled data centres use more water." Whether a site consumes water depends on how it finally rejects heat (evaporative towers or dry coolers), not on whether the chips are cooled by liquid. Warm liquid cooling often makes dry cooling possible.

Misconception

"A heat pipe conducts heat like a copper rod, only better." It works only within its limits (capillary, boiling, sonic and others), depends on orientation, and has a startup temperature. Past the capillary limit the evaporator dries out and its temperature climbs sharply.

Exercises

Exercises

1. Thermal management for computing

Warm-upHuman

A thermal interface material with an area-specific resistance of 0.15 cm²·K/W sits between a 6 cm² die and its lid. The chip dissipates 400 W, spread evenly. What is the temperature drop across the interface, in kelvin?

2. Thermal management for computing

CoreHuman

An accelerator dissipating 1200 W has a thermal resistance of 0.025 K/W from junction to the coolant flowing past it (including hot spots). The coolant warms by 8 K across the cold plate, and the chip sees the mean coolant temperature. Its junction must stay at or below 90 °C. What is the highest coolant inlet temperature allowed, in °C?

3. Thermal management for computing

Warm-upHuman

A rack sends 100 kW of heat into water (cp=4180c_p = 4180 J/(kg·K), density 1000 kg/m³) that warms by 10 K. What flow rate is needed, in litres per minute, to the nearest whole number?

4. Thermal management for computing

CoreComputational

Write two functions for a plate-fin heat sink.

fin_efficiency(h, k, t, H) returns ηf=tanh⁡(mH)/(mH)\eta_f = \tanh(mH)/(mH) with m=2h/(kt)m = \sqrt{2h/(kt)}, for a fin of thickness t and height H (m), conductivity k and heat-transfer coefficient h.

sink_resistance(h, N, H, L, t, W, k, C) returns the resistance (K/W) from base to inlet air of a heat sink with N fins of height H, flow length L and thickness t on a base of width W: R=1/[C(1−e−hA/C)]R = 1/[C(1 - e^{-hA/C})], where hA=h(ηf⋅2NHL+(W−Nt)L)hA = h(\eta_f \cdot 2NHL + (W - Nt)L) and C is the air's heat-capacity rate in W/K.

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

5. Thermal management for computing

CoreComputational

Write chf_zuber(rho_l, rho_v, h_fg, sigma, g=9.81, C=0.131) returning Zuber's critical heat flux qmax′′=C hfg ρv1/2[σg(ρl−ρv)]1/4q''_\text{max} = C\,h_{fg}\,\rho_v^{1/2}[\sigma g(\rho_l - \rho_v)]^{1/4} in W/m². The tests check water and a dielectric liquid at 1 atm, and how the critical heat flux changes in reduced gravity.

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

6. Thermal management for computing

CoreHuman

A laptop's heat pipe works well on a desk, but a user props the laptop vertically with the processor (the evaporator end) at the top and the fan's fins (the condenser) at the bottom. What most likely happens under heavy load?

Choose one

7. Thermal management for computing

CoreHuman

A data-centre operator raises the facility water supply temperature from 20 °C to 40 °C, after confirming the chips stay within their limits. Which consequences are expected? Select all that apply.

Choose all that apply

8. Thermal management for computing

StretchHuman

Which statements about cooling AI racks are accurate as of 2026? Select all that apply.

Choose all that apply

Further reading and resources

  • High-performance heat sinking for VLSI · David Tuckerman and Fabian Pease · 1981IEEE Electron Device Letters 2(5), 126–129. The first microchannel heat sink, etched into silicon, which removed 790 W/cm².
  • Co-designing electronics with microfluidics for more sustainable cooling · Remco van Erp, Reza Soleimanzadeh, Luca Nela, Georgios Kampitsis and Elison Matioli · 2020Nature 585, 211–216. Coolant channels built into the chip itself, from EPFL in Lausanne.
  • On the stability of boiling heat transfer · Novak Zuber · 1958Transactions of the ASME 80, 711–720. The hydrodynamic theory of the critical heat flux.
  • ASHRAE Technical Committee 9.9Publishes the thermal guidelines for data processing environments, including the air and liquid-cooling classes.
  • Open Compute Project, Cooling EnvironmentsOpen specifications for cold plates, coolant distribution units and immersion systems.
  • JEDECIts JESD51 series defines how package thermal resistances are measured and reported.
  • The Green GridThe industry body behind the PUE and WUE metrics.
  • Heat Pipe Science and Technology · Amir Faghri · 2016A comprehensive text on heat pipes, their limits and their design (2nd edition, Global Digital Press).
  • Fundamentals of Heat and Mass Transfer · Theodore Bergman, Adrienne Lavine, Frank Incropera and David DeWittThe standard undergraduate text, with the fin, internal-flow and boiling relations used here.