Digital Signal Processing Reference
In-Depth Information
Power consumed by a homogeneous multiprocessing system is given by
1
CPI
SLE
V
MP
.
≈
∗
∗
∗
P
MP
P
C
(14)
Again, as in the case of superscalar processor, the term considering the CPI of
the single low-end processor
1
CPI
SLE
has been also included, while its energy is
given by
E
MP
=
P
MP
∗
T
MP
.
(15)
When one compares both, it is possible to write
CPI
SLE
E
SHE
E
MP
=
CPI
SLE
issue
+
β
1
CPI
SHE
∗
1
K
issue
∗
∝
∗
P
+
γ
1
CPI
MP
,
=
P
∗
∗
(
α
CPI
SLE
+
β
CPI
SLE
)
∗
(16)
V
SHE
(
∝
+
K
1
issue
∗
β
)
∗
E
SHE
E
MP
=
.
(17)
V
MP
[(
δ
+
P
γ
)]
when there is no ILP or TLP available since there is no power management
energy results considering the same power budget, as it was already done in the
performance model. For this first experiment, we do not consider the communication
overhead for the multiprocessing environment that will be modeled later. In
addition, we only show the energy of 8- and 18-Core Designs, since the conclusions
of these setups are also valid for the rest of the setups.
The high-end single processor organization spends higher energy than the 18-
Core multiprocessor the same amount of energy when considering all levels of
To obey the given power budget, the 8-Core multiprocessor runs 3 times faster
than four-issue superscalar and the 18-Core multiprocessor. Thus, as the 8-Core
Design present 3 times lower execution time than the 4-issue superscalar, the former
spends 3 times less energy. When the parallelism is more exposed the superscalar
approaches to the 8-Core Design, since its execution time decreases. Multipro-
cessors composed of a significant number of cores present worst performance in
no power management techniques are considered (e.g., cores are turned off when
not used), energy consumption of such multiprocessor designs tend to be higher