4. Design Overhead Analysis
As DRAM is competing at a thin profit margin, and the cost
(known as $/bit) is very sensitive to area increase, any change
in DRAM structure should be assessed with area overhead
analysis. In this section, we will conduct a detailed analysis
on DRAM area overhead to justify the practicality of our
Half-DRAM design. Typically, the commodity DRAM is
implemented with three metal layers [23, 28]. This design
can be well leveraged by our Half-DRAM design. Figure 9
presents the circuit design of Half-DRAM. Instead of driving
a local wordline that traverses all 512 bitlines in a single
MAT (Figure 9a), the wordline can be horizontally shifted and
cover a half row from both neighboring MATs, respectively.
Note that the two half-row wordlines share a common row
logic stripe (Figure 9b). In other words, the local wordline
driver becomes bi-directional, and drives the wordline (or
row) at both directions. Obviously, this design does not incur
extra area or routing overhead, given that it only involves a
horizontal shift of the metal wordline and only requires a few
additional metal vias to connect to the driver at circuit level.
Figure 9 also shows the proposed design change of column
select lines (CSLs) routing. In the original DRAM layout as
Figure 9a shows, the CSLs are organized in an order that is
identical between MATs. One CSL connects to four I/O transistors
to select data out. Therefore, each MAT can have only
one ×4 column selected and buffered in the HFFs assigned
to the MAT. In Half-DRAM, in order to select the desired
data within one even or odd DRAM row, the connection of
CSLs to the output of column decoder is mirrored between
even and odd MATs. For example, two CSLs are shown in
the figure with each from the left and right slab, respectively.
Originally, if the output of column decoder is ‘10’, then the
columns on the left half in both MATs are selected due to the
homogeneous connection (in red color in Figure 9a). Instead,
as the CSL connection is mirrored, the same decoding output
4. Design Overhead Analysis
As DRAM is competing at a thin profit margin, and the cost
(known as $/bit) is very sensitive to area increase, any change
in DRAM structure should be assessed with area overhead
analysis. In this section, we will conduct a detailed analysis
on DRAM area overhead to justify the practicality of our
Half-DRAM design. Typically, the commodity DRAM is
implemented with three metal layers [23, 28]. This design
can be well leveraged by our Half-DRAM design. Figure 9
presents the circuit design of Half-DRAM. Instead of driving
a local wordline that traverses all 512 bitlines in a single
MAT (Figure 9a), the wordline can be horizontally shifted and
cover a half row from both neighboring MATs, respectively.
Note that the two half-row wordlines share a common row
logic stripe (Figure 9b). In other words, the local wordline
driver becomes bi-directional, and drives the wordline (or
row) at both directions. Obviously, this design does not incur
extra area or routing overhead, given that it only involves a
horizontal shift of the metal wordline and only requires a few
additional metal vias to connect to the driver at circuit level.
Figure 9 also shows the proposed design change of column
select lines (CSLs) routing. In the original DRAM layout as
Figure 9a shows, the CSLs are organized in an order that is
identical between MATs. One CSL connects to four I/O transistors
to select data out. Therefore, each MAT can have only
one ×4 column selected and buffered in the HFFs assigned
to the MAT. In Half-DRAM, in order to select the desired
data within one even or odd DRAM row, the connection of
CSLs to the output of column decoder is mirrored between
even and odd MATs. For example, two CSLs are shown in
the figure with each from the left and right slab, respectively.
Originally, if the output of column decoder is ‘10’, then the
columns on the left half in both MATs are selected due to the
homogeneous connection (in red color in Figure 9a). Instead,
as the CSL connection is mirrored, the same decoding output
การแปล กรุณารอสักครู่..