Source/Drain Epitaxy Process
Before the replacement-gate module and the silicide layer this batch already built can do anything, the source and drain regions sitting on either side of the channel have to be doped and, in modern logic, mechanically strained, and that is the job of source/drain epitaxy. Rather than simply implanting dopant ions into the existing silicon and annealing, this step etches the source/drain regions out entirely and regrows them from scratch with a different, strained crystal: silicon-germanium (SiGe) for PMOS, under compressive strain, and silicon-carbon (Si:C) or phosphorus-doped silicon (SiP) for NMOS, under tensile strain. The strain does not stay confined to the regrown region; because the new material is epitaxially locked to the surrounding silicon lattice, its strain propagates laterally into the channel directly underneath the gate, squeezing or stretching the channel's crystal lattice and measurably changing how fast holes or electrons move through it. Intel was first to put this into production, at the 90nm node in 2003, with compressively strained SiGe source/drains for PMOS; the company reported roughly a 25% PMOS drive-current improvement from the technique, a result commonly cited in the literature, and TSMC, Samsung, and GlobalFoundries all adopted embedded strained source/drain epitaxy industry-wide in the generations that followed.
PMOS and NMOS need opposite strain, and that is why they use different materials and different recess geometries. Germanium's lattice constant is larger than silicon's, so a SiGe crystal grown inside a silicon recess is squeezed to fit, and that squeeze transmits as compressive stress into the adjacent channel; compressive strain along the channel raises hole mobility, which is exactly the carrier PMOS depends on, so embedded SiGe source/drains became the standard PMOS strain technique. Carbon's lattice constant is smaller than silicon's, so a Si:C crystal is stretched to fit, transmitting tensile stress into the channel instead; tensile strain raises electron mobility, the carrier NMOS depends on, so Si:C (or phosphorus-doped silicon, SiP, used by some manufacturers for a similar tensile effect) became the NMOS counterpart. For PMOS, the recess etch is typically shaped into a distinctive sigma (Σ) or faceted cavity rather than a simple rectangular box, because the pointed facets push the strained SiGe material's stress field closer to and more directly under the channel, extracting more mobility benefit from the same germanium content; NMOS recesses are more often simpler rectangular or less aggressively faceted cavities, since Si:C's tensile effect does not require the same geometric stress concentration to be effective.
The process itself runs in a fixed sequence, interleaved with the gate module rather than happening in isolation. It starts only after gate and spacer formation is in place, because the spacer defines exactly where the source/drain recess etch is allowed to open and protects the gate edge during that etch; an anisotropic recess etch then carves out the source/drain silicon, forming the sigma-shaped cavity for PMOS or a more rectangular cavity for NMOS; a pre-epi clean follows immediately, and this step is unforgiving, since any residual native oxide or etch-induced surface damage left in the cavity will seed crystal defects or block epitaxial nucleation outright, ruining the strain transfer the whole step exists to deliver; selective epitaxial growth by chemical vapor deposition then fills the clean cavity, using a dichlorosilane/germane (SiH2Cl2/GeH4) chemistry with in-situ boron doping for PMOS SiGe, or a silane/methylsilane-type (SiH4/CH3SiH3) chemistry with in-situ phosphorus or carbon doping for NMOS; the growth is selective, meaning it nucleates only on exposed silicon in the cavity and not on the surrounding oxide or spacer, which is what lets it fill precisely where it is needed. An optional thin silicon or silicon-germanium cap layer sometimes follows the main fill, improving surface quality for the silicide step that comes next. Typical SiGe source/drain germanium content runs in the 20-40 atomic percent range; higher germanium content increases lattice mismatch and therefore compressive strain, but pushes closer to the defect-generation limit the pre-epi clean and recess geometry are managing against.
| PMOS: embedded SiGe | NMOS: embedded Si:C (or SiP) | |
|---|---|---|
| Strain type | Compressive | Tensile |
| Lattice mismatch driver | Ge larger than Si, compresses to fit | C smaller than Si, stretches to fit (SiP: P substitution for similar tensile effect) |
| Typical concentration | ~20-40 atomic % Ge | Low atomic % C (small substitutional fraction); SiP uses phosphorus doping level instead |
| Mobility benefit mechanism | Compressive channel strain raises hole mobility | Tensile channel strain raises electron mobility |
| In-situ doping species | Boron (B) | Phosphorus (P) or carbon (C) |
| Recess cavity shape | Sigma (Σ) / faceted, concentrates stress under channel | Simpler rectangular, less geometric concentration needed |
| CVD precursor chemistry | SiH2Cl2 / GeH4 | SiH4 / CH3SiH3-type |
| First production adoption | Intel 90nm, 2003 | Followed within a few generations industry-wide |
| Industry adoption | Intel first; TSMC, Samsung, GlobalFoundries industry-wide since | Same adopters, slightly later generations |
This strain engineering is distinct from, but works alongside, the strain contributions already covered in this batch's gate stack and contact chain. High-k/metal gate formation and the replacement-gate flow set the gate stack's electrical and reliability characteristics, not the channel's mechanical strain state; silicide formation turns the finished source/drain and gate surfaces into an ohmic contact surface, not a strain source. Source/drain epitaxy is the step that actually puts mechanical stress into the channel, and because the recessed cavity has to exist before the dummy-gate or replacement-gate module finishes its own sequence, the two processes are interleaved in the real integration flow rather than fully separate: spacer formation defines both the recess etch boundary for epitaxy and the placeholder gate's footprint, so the two modules share that starting geometry.
Read end to end, this pushes the FEOL-to-BEOL arc this batch has been building one step further back, to the transistor's fundamental channel engineering. Source/drain epitaxy sets the strained channel and doped, strained source/drain regions that everything downstream depends on: the high-k/metal gate replacement-gate module forms around and on top of this same geometry, silicide formation converts the strained source/drain surface this step produces into a contactable surface, and the tungsten and cobalt plug-fill chain completes the wiring on top of all of it. Without the compressive SiGe or tensile Si:C strain established here, the gate stack and contact chain built on top would still function, but the drive-current and mobility gains that have driven planar and FinFET scaling since Intel's 90nm node would not exist.