THE COURSE / M1 THE SWITCH / M1.6
Scaling and its limits
After this lesson you can:
- Say what Dennard scaling promised: smaller, faster, and the same power per area.
- Explain why the supply voltage stopped falling, and what that did to power.
- Explain why a short planar channel leaks, and what the industry did to buy time.
Loading the lesson…
Sources
- R. H. Dennard, "This Week's Citation Classic" commentary on Dennard et al., "Design of ion-implanted MOSFET's with very small physical dimensions," IEEE J. Solid-State Circuits SC-9 (1974), Current Contents (1982). From 1970 an IBM group led by D. L. Critchlow aimed to shrink dimensions from 5 µm to about 1 µm. FETs showed a precipitous drop in the gate voltage needed to switch them once the source–drain spacing fell below about 4 µm, as the depletion regions around source and drain merged and interfered with the gate. Reducing the voltage and raising the substrate doping by a factor α shrinks the depletion regions by α; all dimensions can then be reduced by α, keeping the same field patterns: α times the performance and an α² reduction in power, so the denser circuits can be cooled. Scaled devices are named after the paper.
- C. C. Hu, Modern Semiconductor Devices for Integrated Circuits (Pearson, 2010), chapter 7, "MOSFETs in ICs—Scaling, Leakage, and Other Topics" (free from the author). Gate oxide thinned from 300 nm at the 10 µm technology to 1.2 nm at 65 nm, roughly in proportion to the line width, to raise the on-current and to control Vt roll-off as the channel shortens. Below 1.5 nm, tunnelling leakage limits it: 1.2 nm of SiO₂ leaks 10³ A/cm², 10 A from 1 mm² of gate, enough to drain a phone's battery in minutes; nitrogen cuts it about tenfold. HfO₂ has k ≈ 24, six times SiO₂'s: 6 nm of HfO₂ gives the capacitance of 1 nm of SiO₂ (an EOT of 1 nm) with leakage several orders of magnitude lower; a thin SiO₂ interfacial layer under it; metal gates, two metals for N and P transistors. SiO₂ breakdown about 7 MV/cm for ten years' operation. Below threshold the current falls exponentially, by ten for every S = η × 60 mV at room temperature (η ≥ 1; e.g. 90 mV at η = 1.5): the subthreshold swing. Vt is defined where Ids = 100 nA × W/L (some companies 200 nA); Ioff, at Vgs = 0, = 100 nA × (W/L) × 10^(−Vt/S). An Ioff of 100 nA per transistor across a hundred million transistors is 10 A in standby. Vt can't be made arbitrarily small; S is lowered by a thinner oxide or a wider depletion layer. Vt rolls off as the channel shortens: short-channel transistors leak more. Moore (1965): devices per chip double every 18 to 24 months; each node shrinks features to 70%, halving the area of a circuit, every two to three years. Table 7–1 (history and ITRS projection): 90 nm (2003) Vdd 1.2 V, 65 nm 1.1, 45 nm 1.0, 32 nm 1.0, 22 nm (2013) 0.9 V; strained silicon from 90 nm, high-k/metal gate from 45 nm, immersion lithography at 32 nm, a new structure at 22 nm. Strain: SiGe grown in trenches beside a PFET's gate compresses its channel and raises hole mobility. Vt roll-off: in a short channel the drain lowers the source–channel barrier (drain-induced barrier lowering), so less gate voltage is needed; roll-off ∝ (Vds + 0.4 V) e^(−L/ld), ld ∝ (Toxe Wdep Xj)^(1/3); minimum L several times ld. Even with a thin oxide the drain controls leakage paths below the surface, far from the gate: ultra-thin bodies and multigate transistors remove them. Random dopant fluctuation: the number and place of dopant atoms in a small transistor varies, and with it Vt. Designers get two or three Vts to choose from.