DRAM Buried Word Line
DRAM Buried Word Line: Saddle-Fin Gate Architecture, the GIDL Trade-Off, and Hierarchical Row Activation
A buried word line (BWL) is a $\mathrm{DRAM} \text{access-transistor}$ gate that no longer sits on top of the silicon — it is recessed into a trench etched below the original surface, wrapping around a narrow silicon fin on three sides instead of controlling the channel from one flat face above it. DRAM moved to this geometry because a 1T1C access transistor at DRAM's pitch cannot get a planar gate long enough to stay off reliably; recessing the gate buys channel length without spending any extra footprint, at the cost of a leakage $\text{trade-off}$ that did not exist when the gate sat flat on top of the silicon.
Getting a full logic voltage onto the storage node at all requires overdriving the gate. An NMOS access transistor can only pass a voltage up to $V_{\text{dd}} - V_{\text{th}}$ onto its source before it starts to turn itself off, so writing a full $V_{\text{dd}}$ onto the storage capacitor needs the word line driven above the supply rail entirely:
Every DRAM chip generates this boosted voltage, called $V_{\text{pp}}, \text{on-die}$ with a dedicated charge pump, purely so the access transistor can pass a clean "1" without losing a threshold voltage's worth of charge on the way in.
That same transistor has to shut off just as cleanly, and that is where the recessed geometry creates a new problem. A buried gate's sidewall sits close to, and slightly underlaps, the drain diffusion above it — exactly the geometry that produces $\text{gate-induced}$ drain leakage $(\mathrm{GIDL})$: band-to-band tunneling at the $\text{gate-drain}$ overlap, driven by the electric field there, which gets $\cdot stronger^{\ast}$, not weaker, the more negative the word line is pushed in the off state. Subthreshold leakage through the channel and GIDL at the drain overlap respond to the $\text{off-state}$ bias in opposite directions, which makes the word line's "off" voltage a genuine optimization, not just "as negative as possible":
That minimum is tuned once per design, but the word line still has to physically reach every cell fast enough, and a buried line is a worse conductor than a surface metal one. Recessed polysilicon or tungsten inside a narrow trench carries far more resistance per micron than a metal interconnect line, and the RC delay of driving it from one end scales with the square of its length:
Doubling how many cells one continuous word line serves quadruples its activation delay — an unacceptable trade at DRAM's density — so no production array drives a buried word line end-to-end from a single driver.
All three constraints — the $V_{\text{pp}}$ overdrive, the GIDL-versus-subthreshold minimum, and the hierarchical driver split — are decided together, not separately. Recess depth and gate work function set how much overdrive the transistor needs and how sharp the GIDL corner is; the chosen $V_{kk} \text{back-bias}$ sets how hard the charge pumps and $\text{substrate-bias}$ generators on the die have to work; and the sub-word-line pitch sets how many drivers a given array floorplan needs room for. A DRAM maker that designs the cell, the process recipe, and the $\text{on-die}$ voltage generation together can tune all three against each other; a team trying to license just the transistor geometry inherits none of that $\text{co-design}$, which is the same reason the capacitor and the isolation trench turned out to be moats in their own right — the buried word line is a third structure in the same 1T1C cell where the manufacturing process and the circuit design are inseparable.
Read the buried word line through a $\cdot \text{field-at-the-corner}^{\ast}$ lens rather than a $*"just$ a recessed $gate"*$ lens: every design choice here — how deep to recess it, how negative to drive it off, how many cells one segment gets to serve — is a different point on the same curve in $I_{\text{total}}(V) = I_{\text{subthreshold}}(V) + I_{\mathrm{GIDL}}(V)$, traded off against the $t_{\mathrm{WL}} \propto R_{\mathrm{WLC}_{\mathrm{WLN}}}^2$ delay that decides how finely the array has to be split to activate a row fast enough to matter.