1 INTRODUCTION
As a topic of great current interest, three-dimensional (3D)
integration provides a viable and promising option to address
the well-known memory wall problem [1] in high-performance
computing systems, i.e., by stacking multiple high-capacity
DRAM dies with one or more processor dies and providing
massive inter-die interconnect bandwidth, 3D processor-
DRAM integrated systems can have drastically reduced memory
access latency and increased memory access bandwidth.
This has been well recognized by the computer architecture
community and many recent works [2]–[8] have explored
and well demonstrated the potential of 3D processor-DRAM
integrated computing systems.
Most prior works [2]–[4] on 3D processor-DRAM integration
assumes the 3D stacking of several conventional commodity
2D DRAM dies as main memory. Loh [5] demonstrated
that the performance of 3D processor-DRAM integrated system
can be further improved by using a non-conventional
so-called “true” 3D DRAM design strategy announced by
Tezzaron Corporation (see [9]). The key feature of this “true”
3D DRAM design strategy is to put DRAM cell arrays and
DRAM peripheral circuits on separate dies so that highperformance
logic die(s) can be used to implement DRAM
peripheral circuits, which is claimed being able to improve the
speed and reduce silicon area. Nevertheless, such an aggressive
design strategy demands the pitch of silicon vias (TSVs) to
be comparable to the DRAM wordeline/bitline pitch (e.g.
few hundreds nm or less), which results in significant TSV
fabrication challenges, particularly as the DRAM technology
continues to scale down. Moreover, we note that all the prior