Online proceedings - EDA Publishing Association

Fig. 1. Package cross-section illustrating the spatially resolved heat transfer
modulation concept for a) normal-flow and b) cross-flow architecture.
The hotspot addressing efficiency strongly depends on the
spreading characteristics of a package. The heat-flux
contrast in the cold plate base compared with the source is
reduced in classical packages because of the 720-μm
silicon die, the bottleneck of the thermal interface material
(TIM), and the cooler base plate. For interlayer-cooled 3Dchip
stacks, where coolant picks up heat 10 to 50 μm from
its source, smaller hotspots can be treated [6]. The gain of
hotspot cooling for direct-attach (Fig. 2a) and TIM-attach
(Fig. 2b) packages are investigated and compared.
Fig. 2. Schematic of a (a) direct-attach and (b) TIM-attach package.
(HT: heat transfer)
To maximize the available exergy, the fluid temperature
increase from inlet to outlet of the cold plate (ΔT fout-in ) at a
given maximal junction temperature (T jmax ) and power
dissipation (P el ) needs to be maximized. This translates
according to the sensible heat ( Q ) definition (1) with fluid
density (ρ) and heat capacity (c p ) into the minimization of
the flow rate (V ):
Q
V =
. (1)
ΔT fout
⋅ c ⋅ ρ
−in
p
This parameter is then used as the cost function in the heattransfer
optimization process, representing the inverse of the
exergy accordingly.
Moreover, the pumping power efficiency of the cold plate
is benchmarked. Most publications compare the cold plate
performance isolated from the complete cooling system.
Considering server-rack liquid cooling with many power
sources and parallel coupled cold plates, this is not a
meaningful metric. Additional fluid loop pressure drops due
to secondary fluid-fluid heat exchanger, filters, and fluid
quick-connections all add up to the total pressure drop. The
total system pumping power for low flow rate cold plates
with moderately increased cold plate pressure drop still is
reduced because of the minimization of the total flow rate
and parasitic pressure drop.
7-9 October 2009, Leuven, Belgium
III. NORMAL-FLOW COLD PLATE STUDY
For this test case, experimentally defined unit cell flow rate
( V ) to heat transfer characteristics (h ) of a
n , m
n,m
commercially available high performance normal-flow cold
plate is used [7] (Fig. 3) (Mikros Technologies). It is
hn
m k
V ,
n, m
= g ⋅( ) ⋅ An
, m
h
, (2)
0
with coefficients g = 4.97 × 10 -11 L/min/cm 2 , k = 1.939 and
h 0 = 1 W/(m 2 *K) for the unit-cell area A. All unit cells are
coupled in parallel to the manifold. Therefore the cell with
the highest heat-flux need defines the pressure drop from
inlet to outlet, and not the size of the cold plate area as in
cross-flow heat exchange. The cold plate flow rate (V ) is
computed by adding the unit-cell flow rates
. (3)
V
= ∑
V n , m
n,
m
The server-rack cooling-loop flow rate ( V
system
) and
pressure drop (Δp loop ) depend on the number of cold plates
(n) coupled in parallel and the coefficient of flow (k v
[L/min]) of the cooling loop:
V
system
= n ⋅V
V
system
and Δp
( )
2
loop
= ⋅ p0
, (4)
kv
with p 0 = 1 bar. For one server rack, we assume 56 cold
plates (Fig.3, green curve). The total system pressure drop is
the server-cooling loop plus the cold-plate pressure drop
Δp system = Δp cp + Δp loop .
Fig. 3. Heat transfer coefficient (h eff) and pressure drop characteristics of
the uniform cold plate at normalized volumetric flow rates (Δp cp) and a
typical server fluid loop pressure drop (Δp loop) attached to 56 parallel cold
plates.
A simplified power map of a high-performance processor
was chosen as model input (Fig. 4). The peak heat flux is
three to four times higher than the average power density of
50 W/cm 2 . Total power dissipation is 120W and the hotspot
area fill-factor is 10%.
©EDA Publishing/THERMINIC 2009 151
ISBN: 978-2-35500-010-2