High-K Metal Gate (HKMG) Process
The gate stack that the rest of this batch's contact chain sits on top of is not poly-Si on SiO2 anymore, and the reason it changed is a hard physical wall silicon dioxide hit on its own: gate oxide scaling ran out of room once the physical thickness dropped to roughly 1 nm, the point at which direct quantum-mechanical tunneling through the oxide turns gate leakage current into a runaway power and reliability problem rather than a second-order effect. Silicon dioxide's dielectric constant, about 3.9, ties physical thickness directly to electrical thickness (equivalent oxide thickness, EOT): to keep shrinking EOT and keep transistor drive current scaling with the node, the physical oxide had to keep shrinking too, until there was no more margin between "thin enough for the capacitance the device needs" and "thin enough that electrons tunnel straight through it." Hafnium oxide (HfO2), with a dielectric constant around 20-25, breaks that link: a physically thicker HfO2 film gives the same EOT as an ultra-thin SiO2 film while presenting a far higher tunneling barrier, cutting gate leakage current by orders of magnitude at equivalent capacitance. Intel's 45nm node in 2007 was the first production node to ship this high-k dielectric paired with a metal gate electrode in place of the legacy poly-Si/SiON gate stack, and the pairing was not optional: a high-k dielectric against a poly-Si gate suffers Fermi-level pinning and mobility-degrading phonon scattering that erase most of the leakage benefit, so a metal gate electrode has to come with it to set a usable threshold voltage and preserve carrier mobility.
Two integration schemes exist for building this high-k/metal gate stack, and only one of them survived contact with high-volume manufacturing. Gate-first integration deposits the high-k dielectric and the final metal gate electrode early in the flow, before the source/drain implant anneal that activates dopants at temperatures well above 1000 C; the metal gate and high-k film then have to survive that anneal intact, and in practice they do not survive it well; threshold voltage drifts, work-function metals interdiffuse or react with the high-k layer, and the device-to-device variability this introduces made gate-first difficult to control at production yield. Gate-last integration, also called the replacement-gate flow, solves this by sequencing the anneal around a sacrificial placeholder instead: a dummy poly-Si gate goes in first, survives the full source/drain anneal exactly the way the legacy poly-Si process always did, and only after that thermal budget is spent does the dummy gate get removed and replaced with the real high-k/metal gate stack. This is the approach Intel, TSMC, Samsung, and GlobalFoundries have all converged on industry-wide since roughly the 32/28nm generation, and gate-first is now largely a historical footnote rather than a competing production option.
The order of that flow is not arbitrary, it is the entire reason gate-last exists: every step before dummy poly removal is there specifically to let the wafer absorb the source/drain anneal's thermal budget on a material that can take it, so that the high-k film and the work-function metal only ever see a comparatively mild thermal history after they are deposited. Removing the dummy gate late also protects a second thing gate-first integration could not protect: the work-function metal's as-deposited composition and interface chemistry, which set the threshold voltage, stay fixed from the point they are deposited onward instead of drifting through a high-temperature anneal the way a gate-first stack's metal would. The thin interfacial SiO2 layer deposited just before the HfO2 ALD step looks like a step backward, since the whole point of high-k was removing SiO2 from the stack, but it exists to give the ALD-grown HfO2 a lower-defect-density surface to nucleate on; the EOT contribution of that interfacial layer is small enough that the overall EOT benefit of high-k is preserved while interface-trap density, and the mobility degradation that comes with it, drops substantially compared with growing HfO2 directly on bare silicon.
| SiON/poly-Si legacy gate | HfO2/metal HKMG gate | Gate-first integration | Gate-last (replacement-gate) integration | |
|---|---|---|---|---|
| Dielectric constant (k) | ~3.9-5 | ~20-25 | ~20-25 (same material) | ~20-25 (same material) |
| Gate leakage at scaled EOT | Approaches direct-tunneling wall below ~1nm physical thickness | Orders of magnitude lower at equal EOT | Same material benefit, but degraded by anneal exposure | Same material benefit, fully preserved |
| Thermal budget compatibility | Poly-Si tolerates full S/D anneal natively | N/A (dielectric alone, pairs with either integration) | Metal/high-k exposed to full S/D anneal; threshold voltage drift and interdiffusion risk | Metal/high-k deposited after anneal; thermal exposure minimized |
| Work-function tuning flexibility | Poly-Si work function limited and Fermi-pinned against high-k | N/A | Must set final work function before surviving the anneal | Work-function metal deposited last, composition stays as-set |
| Process complexity / cost | Lower, fewer unique steps | N/A | Lower step count than gate-last | Higher: extra ILD0 CMP, dummy removal, and gate CMP steps |
| First production adoption | Multiple decades, through ~65nm | Intel 45nm, 2007 | Explored early HKMG era, mostly abandoned | Dominant since ~32/28nm; Intel, TSMC, Samsung, GlobalFoundries |
| Industry outcome | Superseded once tunneling leakage became unmanageable | Enables continued EOT scaling | Largely abandoned for thermal stability and variability reasons | Industry-wide standard integration scheme |
High-k's dielectric constant advantage over SiO2 is the whole leakage story in one number, and the comparison table above is really just that number propagating through two different integration choices. A SiO2 gate oxide scaled to the EOT modern transistors need has to be physically thin enough that electrons tunnel through it at a rate that turns gate leakage into a chip-level power problem; HfO2 at five to six times the dielectric constant can sit at several nanometers of physical thickness, five to six times thicker than an equivalent SiO2 film, while delivering the identical electrical capacitance, and that physical margin is what suppresses direct tunneling by orders of magnitude. That same material improvement is available whether it is integrated gate-first or gate-last, but only gate-last actually delivers it at volume, because gate-first's biggest cost is not the dielectric at all, it is what the high-temperature anneal does to the work-function metal and the high-k film's interface quality once both are locked in before that anneal runs.
This gate module is the piece this batch had been assuming existed every time it talked about contacts: gate+spacer formation complete, named explicitly as the starting point for silicide-formation-process, is the output of this replacement-gate flow, not a given. The dummy-gate removal, high-k ALD deposition, and metal gate fill described here are what actually produce a gate stack with the electrical characteristics modern devices need, and only once that stack is planarized and complete does the silicide layer form on the source/drain and gate to give the tungsten-plug-contact-fill-process and cobalt-contact-fill-process chain an ohmic surface to land on. Read end to end, the arc runs from gate formation through contact metallization in one continuous story: HfO2/metal gate stack formation solves the leakage crisis that killed scaled SiO2, the replacement-gate sequence is the only integration scheme that delivers that solution without the thermal-budget penalty gate-first could not avoid, silicide formation turns the finished gate and source/drain surfaces into something a metal plug can actually contact, and the tungsten and cobalt plug-fill processes complete the wiring on top of all of it.