backside power delivery
**Backside power delivery network (BSPDN)** is the integration architecture that moves VDD and VSS rails to the wafer backside — delivering power through nano-TSVs directly up into the transistors from below, while the frontside metal stack carries only signals. This decouples power and signal routing for the first time in 50 years of planar CMOS, cuts IR-drop by 30–50%, and recovers 1–2 routing tracks per standard cell that would otherwise be consumed by buried power rails. Intel's PowerVia (Intel 20A, 2024) is the industry's first production BSPDN; TSMC's N2P (2026) and Samsung's SF2P follow.
**Why frontside power delivery hit a wall.** In a conventional flow, both signal wires and power rails share the same frontside BEOL (back-end-of-line) metal stack. At sub-3 nm nodes the minimum metal pitch is below 21 nm — so narrow that resistance per unit length climbs steeply (Fuchs–Sondheimer size effects). The power rails must be fat to carry amps, but there's no room for fat wires on a 21 nm pitch. Designers are forced to bury thinner rails in the M0/M1 layers (buried power rail, BPR), which partially solves routing congestion but still can't deliver enough current without excessive voltage drop across the die.
**BSPDN — the architecture.** After frontside fabrication is complete (transistors + full BEOL signal stack), the wafer is bonded face-down to a carrier, thinned from the backside to ~300–500 nm (exposing the shallow-trench isolation), and then a 2–4 layer thick backside metal stack is fabricated for power distribution. Nano-through-silicon vias (nTSVs, diameter 30–80 nm) connect the backside power mesh directly to the source/drain contacts of the transistors below.
$$R_{\text{PDN}} = R_{\text{package}} + R_{\text{µbump}} + R_{\text{backside mesh}} + R_{\text{nTSV}} + R_{\text{contact}}$$
Because the backside mesh uses wider pitches (e.g. 100–200 nm) than the frontside signal layers, each backside metal line has much lower sheet resistance — typically 3–5× lower than an equivalent frontside BPR. Combined with the short vertical path through the nTSV, total PDN resistance drops by 30–50%.
**IR-drop improvement — the quantitative case.** For a 100 W die at $V_{\text{DD}}$ = 0.7 V drawing 143 A peak:
| Metric | Frontside BPR (N3-class) | BSPDN (N2-class) | Improvement |
|---|---|---|---|
| Power rail pitch | 48 nm (shared M0) | 100–200 nm (dedicated backside) | ~3× wider |
| Rail sheet resistance | ~120 Ω/□ (Cu, 20 nm line) | ~30 Ω/□ (Cu, 80 nm line) | 4× lower |
| PDN effective resistance | ~8–10 mΩ | ~4–5 mΩ | ~50% reduction |
| IR-drop (worst-case) | ~70–100 mV (10–14% $V_{\text{DD}}$) | ~35–50 mV (5–7% $V_{\text{DD}}$) | ~50% lower |
| Routing tracks recovered | 0 (BPR in M0) | 1–2 tracks per cell | 15–20% signal density gain |
| Cell height reduction | 5T → 5T (limited by power) | 5T → 4.5T or 4T possible | 10–20% area |
| $L \cdot di/dt$ transient noise | Severe (long frontside path) | Reduced (short nTSV + thick mesh) | ~30% lower |
**The fabrication flow — five new modules.** BSPDN adds significant process complexity beyond standard FinFET/GAA:
1. **nTSV formation.** Deep (200–500 nm) vias etched through silicon and STI from the frontside, landing on a backside metal contact pad. Aspect ratio 5:1 to 10:1 — filled with tungsten or ruthenium liner + copper. Must be placed before transistor gate formation to avoid thermal budget conflicts.
2. **Wafer thinning.** After full frontside processing, the wafer is bonded face-down to a temporary carrier (oxide fusion bond at ~200°C). The silicon substrate is ground and CMP-polished from the backside to ~300–500 nm, exposing the STI oxide and nTSV tips. Uniformity requirement: ±5 nm across 300 mm.
3. **Backside metallization.** 2–4 metal layers deposited on the thinned backside: thick power rails (80–200 nm lines), with low-k dielectric between layers. These layers are fabricated at ≤400°C to protect the frontside devices and bond interface.
4. **Backside patterning.** Alignment to frontside features through the ~300 nm silicon requires infrared (IR) through-wafer alignment or embedded alignment marks — a new challenge for lithography.
5. **Carrier debond and packaging.** The temporary carrier is removed (laser or thermal debond), and the die is packaged with C4 bumps or hybrid bonds connecting to the backside power mesh.
**Thermal implications.** Thinning the silicon to 300–500 nm dramatically reduces the lateral heat-spreading capability of the substrate. In a conventional die, the ~750 µm bulk silicon spreads hotspot power over a wide area before it reaches the thermal interface. With BSPDN, vertical thermal resistance through the thin remaining silicon is low, but lateral spreading is severely curtailed — making localized hotspots worse. Mitigation strategies: thermally-conductive filler in nTSVs, backside metal rails as supplemental heat conduits, and package-level cooling solutions that couple directly to the backside metal.
**Design-rule impact.** BSPDN fundamentally changes PnR (place-and-route) constraints: power is no longer a routing resource that competes with signals. The standard-cell library can eliminate frontside power rails entirely, recovering tracks for signal routing. This enables:
- **Smaller cell heights** (4-track or 4.5-track cells become practical without power-rail area overhead)
- **Higher signal routability** (15–20% more routing resources in M1–M3)
- **Simpler clock-tree synthesis** (less congestion from power blockages)
- **Better timing closure** (lower IR-drop reduces timing-margin guardbands by 20–30 ps)
```svg
```
**The competitive landscape.** Intel demonstrated PowerVia on test chips in 2023 and ships it in production with Intel 20A (2024), making it the first company to deliver BSPDN at scale. The measured results: 6% frequency uplift at iso-power, or 30% power reduction at iso-frequency, on ring-oscillator test structures — attributed entirely to the lower PDN impedance. TSMC's N2P (2025–2026 ramp) adds BSPDN to the nanosheet GAA baseline, and Samsung's SF2P follows in a similar timeframe. IMEC's research roadmap extends BSPDN to direct backside contacts (no nTSV) using buried interconnect formed before transistor processing.
**What BSPDN means for AI accelerator design.** Modern AI chips (H100, MI300, Gaudi) operate at 700–1000 W TDP, drawing 1000+ amperes at sub-0.8V supply. IR-drop is the single largest timing-margin consumer at these power levels. A 50% IR-drop reduction from BSPDN translates directly into either: (a) 5–10% higher $f_{\text{max}}$ at the same voltage, (b) ~15% lower power at the same frequency (from running at lower $V_{\text{DD}}$ guard-band), or (c) supporting larger die sizes before IR-drop forces voltage domain partitioning. This is why every leading-edge AI accelerator design team (Nvidia, AMD, Google, Meta) is designing for BSPDN-enabled nodes.