DEV Community

plasmon
plasmon

Posted on

They Routed Power Through the Back of the Chip and 30% IR Drop Vanished

They Routed Power Through the Back of the Chip and 30% IR Drop Vanished

Every semiconductor chip has a front and a back. Transistors and wiring are built on the front side, and power delivery shares that same front surface with signal routing. It has been this way for over 60 years.

In 2026, Intel broke this convention. They implemented BSPDN (Backside Power Delivery Network) — branded PowerVia — in their 18A process node, shipping it in the Panther Lake processor as a mass-produced product.

The result: 30% IR drop reduction, 6% frequency uplift, and 5-10% standard cell utilization improvement (Intel's CES 2026 announced figures).

If the front side is congested, use the back. It sounds obvious in hindsight, but getting there required decades of process engineering. This article breaks down the physics behind BSPDN, the Intel vs TSMC vs Samsung race, and what it means for AI chips.


Why Power Delivery Became the Bottleneck

Front-Side Congestion

Look at a cross-section of any leading-edge chip: 10+ metal layers are stacked on top of the transistor layer.

Conventional chip cross-section (front-side only):

  Surface
  ─────────────────────────
  M12  Power rails (VDD/VSS) ← thick wires
  M11  Power + signal         ← shared
  M10  Signal routing
  ...
  M3   Signal routing         ← thin
  M2   Signal routing         ← thinner
  M1   Local interconnect     ← thinnest
  ─────────────────────────
  Transistor layer (FinFET/GAA)
  ─────────────────────────
  Silicon substrate
  ─────────────────────────
  Backside (nothing here)
Enter fullscreen mode Exit fullscreen mode

The problem: power lines and signal lines share the same metal layers. Power rails demand thick wires, eating into the space available for signal routing.

Of the 12 metal layers, 3-4 (M10-M12) are consumed by power rails. The remaining 8-9 (M1-M9) carry signals, but power vias — vertical connections that punch through every layer — eat into signal space too.

It gets worse with every node shrink:

Node Signal Congestion IR Drop
5nm Moderate Manageable
3nm High Design problem
2nm Critical Design-limiting

At the 2nm generation, signal routing density physically runs out. Power infrastructure consumes too much space to pack in more logic.

IR Drop: A Power Quality Problem

IR drop is the voltage loss caused by resistance in the power delivery network. The longer the path from the power pin to the transistor, the more voltage you lose.

Power delivery path (conventional):
  Package → bonding pad → M12 → M11 → ... → M1 → transistor

  Each metal layer connected vertically by vias
  Horizontal runs add resistance too
  Long path → high resistance → large voltage drop

  The center of the chip suffers the most:
    Pads are at the periphery → long distance to reach center
    → Center transistors operate at lower voltage than edges
    → Must raise global voltage to maintain margin
    → Increased power consumption
Enter fullscreen mode Exit fullscreen mode

The Physics of BSPDN: Power From the Back

BSPDN is built on a simple idea: move the power wiring from the front to the back of the chip.

BSPDN chip cross-section:

  Surface
  ─────────────────────────
  M12  Signal routing  ← power lines gone, space freed
  M11  Signal routing  ← all layers available for signals
  M10  Signal routing
  ...
  M1   Local interconnect
  ─────────────────────────
  Transistor layer (GAA: gate wraps all around the channel)
  ─────────────────────────
  nTSV (nano-TSV)  ← vertical power feed from backside
  ─────────────────────────
  Backside power grid (VDD/VSS)  ← thick wires, plenty of room
  ─────────────────────────
  Backside
Enter fullscreen mode Exit fullscreen mode

How It's Built

1. Flip the wafer
   Build transistors + signal wiring on the front side,
   then thin the wafer and flip it over.

2. Nano-TSV (nTSV) formation
   Etch tiny vias through the silicon substrate
   reaching down to the transistor contact layer.
   Power is delivered directly to the transistors.

3. Backside metallization
   Form a VDD/VSS power grid on the backside.
   Wires can be thicker than front-side (no signal contention).
   IR drop is drastically reduced.
Enter fullscreen mode Exit fullscreen mode

Quantified Gains (Intel 18A PowerVia)

Metric Conventional (front-side power) BSPDN Improvement
IR drop Baseline -30% Better voltage stability
Clock frequency Baseline +6% Higher speed at same power
Standard cell utilization Baseline +5-10% Power vias removed, more room for logic

A 6% frequency gain may look modest in isolation, but this is achieved without shrinking the process node. Normally, a 6% clock uplift is what you get from migrating to the next node. BSPDN alone delivers roughly half a node's worth of performance — for free, in architectural terms.


Intel vs TSMC vs Samsung: The BSPDN Race

Intel 18A: First Mover

Intel 18A PowerVia:
  - Mass production started early 2026 (Panther Lake)
  - Officially announced at CES 2026
  - Combines RibbonFET (GAA) + PowerVia
  - Foundry customers: Microsoft, Amazon
  - Yield: 60-65% (early 2026; recent reports suggest 65-75%). Target: 70-80%

  Technical details:
  - nTSV connects to transistor contact layer
  - Dedicated coarse-pitch power grid on backside
  - Complete separation from signal routing
Enter fullscreen mode Exit fullscreen mode

TSMC: Late but Ambitious

TSMC Super Power Rail (A16, H2 2026 target):
  - TSMC's version of BSPDN
  - Claims more advanced than Intel PowerVia
  - Power connects directly to transistor source/drain
  - N2 (2nm) ships without BSPDN → A16 introduces it

  TSMC's strategy:
  - Ship N2 first (no BSPDN, but GAA)
  - Add BSPDN in A16 (improved N2)
  - Offer customers a clean N2 → A16 migration path
Enter fullscreen mode Exit fullscreen mode

Samsung is also planning BSPDN introduction with their SF2Z node in 2027, but specific technical details remain undisclosed.

The Competitive Landscape

Foundry Node BSPDN Approach Status Risk
Intel 18A PowerVia (nTSV → contact layer) In production (early 2026) Yield improvement is key to profitability
TSMC A16 Super Power Rail (direct source/drain) H2 2026 target No shipping products yet. Yield unknown
Samsung SF2Z Undisclosed 2027 target 3nm yield issues still lingering

On BSPDN specifically, Intel leads TSMC by roughly a year — the first time in nearly a decade that Intel has been ahead on a major process technology. It is not overall process leadership, but BSPDN is the lifeline of Intel's foundry business.


Three Reasons BSPDN Matters for AI Chips

1. Power Efficiency: Better J/token

AI chip power consumption keeps climbing. The H100 draws 700W, the B200 hits 1000W (1200W for the full-spec variant). Data center electricity costs are starting to dominate inference economics.

BSPDN impact on AI chip power:
  30% IR drop reduction → smaller voltage margin → effective power savings

  Example: hypothetical A100 with BSPDN
    Current: 400W (~15% of which is IR drop margin)
    BSPDN: 30% reduction in IR drop margin → ~18W savings
    Inference cost: ~4.5% improvement
Enter fullscreen mode Exit fullscreen mode

4.5% sounds small until you run the numbers at data center scale — it translates to millions of dollars annually.

2. Logic Density: Packing More Tensor Cores

With power vias removed from the front side, standard cell utilization improves 5-10%. That freed space can be filled with additional compute logic:

  • Extra tensor cores ��� higher inference throughput
  • Larger L2 cache → on-chip KV cache for LLM serving
  • Dedicated PIM units → alleviate memory bandwidth bottleneck

3. Better Thermal Management

Moving power wiring to the backside reduces metal density on the front side. Dense metal creates localized hot spots.

Conventional: Power + signal on front → high metal density → heat traps
BSPDN: Power on backside → lower front-side metal density → better heat spreading
        + shorter cooling path from backside

AI chips have a specific problem:
  Tensor cores run dense computation → localized heating
  BSPDN: shorter power path → less Joule heating generated in the first place
Enter fullscreen mode Exit fullscreen mode

The Cost of BSPDN: The Backside Has Walls Too

BSPDN is not free.

Additional Process Steps

Conventional process:
  Front-side lithography + etch + metallization
  → Done

BSPDN process:
  Front-side lithography + etch + signal metallization
  → Wafer thinning (backgrinding)
  → Wafer bonding (attach to carrier wafer)
  → nTSV formation (backside drilling)
  → Backside metallization
  → Additional lithography
  → Done

Additional steps: 5-8
Enter fullscreen mode Exit fullscreen mode

Cost Implications

Additional Step Est. Cost Premium
Wafer thinning (backgrind + CMP) +5%
Carrier wafer bonding +3%
nTSV lithography + etch +8%
Backside metallization +5%
Backside patterning +4%

The total wafer cost increase is substantial, and yield loss risk compounds it. In return: +6% frequency, +5-10% cell utilization, ~4.5% power efficiency. For high-performance AI chips, the gains absorb the cost premium. For cost-sensitive mobile or IoT chips, BSPDN may be overinvestment.


When Does BSPDN Reach Consumer GPUs?

The timeline for BSPDN in consumer GPUs (GeForce, Radeon) is unclear.

NVIDIA:
  Current (Blackwell): No BSPDN
  Next gen Rubin (R100): TSMC N3P. No BSPDN
  Rubin Ultra: TSMC N2 (2027). BSPDN adoption uncertain
  Post-Rubin Ultra: A16 (with BSPDN) is a possibility

AMD:
  Current (RDNA 4): No BSPDN
  Next gen: Depends on TSMC N2 or A16

Consumer reality:
  2026-2027: No BSPDN (N2 generation)
  2028+: A16-based products may appear
  Cost premium will be passed through to GPU prices
Enter fullscreen mode Exit fullscreen mode

As an RTX 4060 8GB user, the tangible benefits of BSPDN probably won't arrive until the generation after next — a 2028-2029 GPU purchase, most likely. However, data center GPUs (successors to H100/B200) may adopt BSPDN as early as 2027. If you use inference APIs, you'll benefit from BSPDN indirectly.


The Backside Was Empty All Along

The essence of BSPDN is a deceptively simple insight: the backside of the chip was unused. Use it.

For 60 years, the back of a chip was a featureless plane. By relocating power infrastructure to that space:

  • Front-side routing congestion is resolved
  • The power delivery path gets shorter, reducing IR drop
  • Frequency goes up, logic density increases

Intel leads, TSMC follows, Samsung waits in the wings. 2026 will be recorded as the first year BSPDN entered mass production in semiconductor history.

Of the three walls facing the 2nm generation — thermal, power, and scaling — BSPDN partially breaks through the power wall. The other walls will need their own innovations, but the paradigm shift of using the backside hints at where semiconductor design is headed.

There was space left after all. You just had to look at it from the other side.


References

  1. Intel. "Panther Lake: Intel 18A with PowerVia and RibbonFET" (CES 2026)
  2. fiisual. "What Is Backside Power Delivery Network (BSPDN)?" (2026) (industry analysis blog)
  3. TrendForce. "Intel's Clearwater Forest Unveils 18A Backside Power" (2025)
  4. "Backside Power Delivery: A Radical Shift in Chip Architecture" (2026)
  5. Lam Research. "Transistor Channel Stress in Backside Power Delivery Networks"
  6. TSMC. A16 Super Power Rail (H2 2026 planned)

Top comments (0)