tivations in the rolling window tFAW as shown in Figure 7b.
Half-DRAM is promising in that it can relax these power constraints
(manifested as timing constraints), resulting from its
design to activate fewer bitlines each time and thus consume
less activation power. Specifically, the four-activation window
constraint can be relaxed since now eight half-row activations
are permitted in any tFAW window. Such improvement can
help the memory with Restrict-ClosePage applied since the
performance of such memory is mainly limited by tFAW [24].
In addition, Half-Bank level parallelism is introduced in
Half-DRAM to further improve the memory parallelism. Different
from the basic design, two half bank latches are utilized
to decouple Odd and Even groups from each other. As a result,
sub-array level parallelism [10], which uses sub-arrays as independent
memory operation corresponder, can be integrated
into Half-DRAM seamlessly. As shown in Figure 7b, two
half-rows in different sub-arrays can be activated without data
path contentions as long as they belong to different half-rows,
which effectively doubles the memory parallelism.
In this work, two Half-DRAM schemes are proposed as
shown in Figure 7a with the baseline illustrated at the leftmost.
The intermediate Half-DRAM that only permits one
half-row activation is presented in the middle (Half-DRAM-
1Row), meaning that no further activation can be issued even if
the next activation goes to the inactive half-row. Alternatively,
the Half-DRAM that allows any two half-rows to be active is
given at the rightmost side of the figure (Half-DRAM-2Row).
In fact, Half-DRAM-1Row can be easily extended to Half-
DRAM-1Row-Demand, in which the other half of the same row
can be activated immediately as long as no precharge is issued.
Note that we retain the same assumptions that no more than
eight half-row activations can be issued in the tFAW rolling
window and two continuous half-row activations must comply
with the tRRD constraint whenever they go to different rows.