analysis finds high-k, recessed transistors in Samsung's SDRAM
90-nm era truly came of age in 2004, according to a presentation at the
recent Advanced Semiconductor Manufacturing Conference (ASMC) in Munich.
Dick James, Chipworks' senior technology analyst, showed how device specifications
from Intel, IBM, AMD, Fujitsu, TI, and Sony/Toshiba as well as fabless-customer
designs manufactured by UMC, TSMC, and Fujitsu compared with the International
Technology Roadmap for Semiconductors (ITRS) and the respective companies'
structural analyses found that none of the devices matched or exceeded
the metal-1 half-pitch or the minimum gate-length dimensions for the 90-nm
node established in the roadmap's 2003 edition. As James pointed out during
his presentation, "In practice, a node is a consensus of what the
leading-edge manufacturers actually do in their chips, not what's defined
in the ITRS." While the half-pitches analyzed were found
to be both larger and smaller than the companies' published data, all
of the samples examined had smaller gate lengths (ranging from 40 to 55
nm) than the chipmakers' publicly shared numbers. The 90-nm analyses also
confirmed the introduction into production of strained silicon using the
of James's reverse-engineering studies barely made it into his ASMC presentation:
Samsung's 512-Mb DDR SDRAM chip, which debuted at the end of 2004 as the
industry's first mass-produced 90-nm device of its kind. "In this
chip," he notes, "Samsung has introduced some significant process
innovations into mass production: the first recessed transistor in a DRAM,
the first use of an aluminum/hafnium oxide high-k dielectric, and likely
the first mass usage of atomic layer deposition (ALD)." In a MICRO
exclusive, James details what he and his colleagues discovered during
their analyses of the Korean chipmaker's innovative device.
device has a conventional stacked capacitor-over-bitline (COB) architecture
for the memory arrays (see Figure 1). It is fabricated using three levels
of metal and six polysilicon layers. The part features 70-nm-wide gate
polysilicon for the recessed-channel transistors in the DRAM array, and
the minimum-pitch metal 1 is 180 nm, making it a true 90-nm product according
to the ITRS definition.
device includes significant technical changes, such as a high-k aluminum/hafnium
oxide dielectric layer in the storage cell and recessed word-line transistors
in the array. These changes. . .were originally targeted for the 80-nm
node, but it seems that Samsung is happy enough with the technology to
bring it forward into its 90-nm product.
die size is a compact 68 mm2. Figure 2 shows a cross section
taken in the plane of the bitline and offers a clear image of the recessed
channel array transistors (RCATs) used for the word-line transistors in
this part. The concept of the RCAT is to increase Leff by recessing the
channel from the silicon surface, thereby reducing leakage and enhancing
data storage characteristics. The physical gate length in the surface
plane is ~70 nm, but the effective channel length is ~350 nm (which compares
with the ~90-nm channel length of a 130-nm SDRAM part).
literature discloses the use of 193-nm [ArF] lithography with a polysilicon
hard mask to define the trenches, followed by a steam oxidation to form
the gate oxide. Close examination of the image shows that the gate-oxide
thickness varies from ~7 nm at the trench bottom to ~14 nm on the trench
walls, presumably because of the different oxidation rates of the different
crystal orientations. The company claims that the RCAT structure reduces
array junction leakage by more than an order of magnitude. There also
seems to be a thin nitride layer in the shallow-trench isolation, again
possibly to reduce leakage by reducing strain in the substrate.
on to the capacitor structure, Figure 3 is a TEM cross-sectional image
of the tops of a pair of the cylindrical capacitors in the memory array.
The "tuning fork" structures are the tops of two storage nodes,
and the microgranular texture around and between them is actually the
titanium nitride (TiN) layer used as the first layer of the common plate
in the array. Polysilicon is then deposited on this TiN layer to fill
in the gaps between the storage nodes and complete the capacitor.
4 is a horizontal TEM section through the array, clearly showing the cylindrical
nature of the storage nodes. The annular dark circles are again the TiN
layer on the inner and outer surfaces of the tuning forks. A notable feature
at this scale is the differing distances between adjacent capacitors,
created when the sacrificial oxide (used for the formation of the cylinders)
is removed. Figures 5 and 6 provide closer looks at the capacitors, where
the dielectric layers on the inside and outside surfaces of the capacitors
can be seen.
5 and 6
feature that surprised us was the variable crystalline nature of the silicon
in the top capacitor plate. In almost all cases, the silicon inside the
cylinders is amorphous, while outside the cylinders it is sometimes polycrystalline
and sometimes amorphous. Figure 6 shows the bilayer nature of the capacitor
dielectric—the aluminum oxide has been deposited first, then there
is a thinner, darker line which is the hafnium oxide. These layers—and
the TiN—have been formed by ALD, probably one of the first high-volume
applications of this technique."—TC