The Real Reason Chips Fail at 3nm: It Is the Wire, Not the Transistor
Archi3 min read
Contents
The Real Reason Chips Fail at 3nm: It Is the Wire, Not the Transistor#
Every few months a new node arrives with a press release about transistor density, and every few months the yields come in lower than the briefing suggested. The usual explanation points at the gate: leakage, variability, the fin getting too thin. That story is convenient because it is partly true. It also misses the harder problem. At 3nm and below, the limiting element in most failing designs is the interconnect, the copper and barrier metal that has to carry current between billions of transistors that now switch faster than the wire can feed them.
What got worse#
Wire resistance has always risen as dimensions shrink, because a narrower line has less cross section for electrons. What changed is the barrier. Copper needs a liner around it, typically tantalum nitride, to stop it diffusing into the surrounding oxide. At 7nm that liner was maybe a fifth of the wire width. At 3nm it can consume more than half. The remaining copper cross section is small enough that resistance climbs steeply, and capacitance, the other half of the delay product, has not fallen to match.
Signals do not care about the node name. They care about RC delay, and RC delay per unit length has been getting worse for several generations. A designer working on a 3nm block can spend months tuning a clock tree only to find the last 200 microns of routing dominate the timing report.
Why the transistor gets the blame#
Foundries publish transistor curves because those are clean to measure on a test structure. Wires are messier. Their behavior depends on what runs next to them, on the pattern density of the layer, on whether the neighboring net is switching in the same direction. A datasheet number for wire resistance is an average that few real routes ever see.
This asymmetry shapes the conversation. When a chip misses its frequency target, the post mortem tends to start with the device because the device model is trusted. Interconnect models are approximations built from test patterns, and the gap between model and silicon is where schedules die.
The workarounds and their price#
Several fixes are in production. Cobalt and ruthenium liners, air gaps between wires, and taller, thinner lines all help. Each buys a few percentage points of resistance and costs process complexity. Taller lines mean deeper etches and harder gap fill. Air gaps mean mechanical fragility and new failure modes in packaging.
Backside power delivery is the largest change in the pipeline. Moving the power grid to the underside of the wafer frees routing resources on the front and lets signal wires be wider for the same pitch. It is a genuine improvement, and it is also a manufacturing step change that will take years to mature. Expect early versions to ship on a narrow set of high-value parts, not across a product line.
The design side is not blameless#
Some of the wire problem is self-inflicted. Teams stack more intellectual property onto a single die because package-level integration has its own cost and risk. That raises the average route length between blocks. A design partitioned across two chiplets linked by a short die-to-die interface often beats a monolithic version on timing, even after paying the packaging penalty.
There is a habit of using the newest node for anything that needs a marketing number, including blocks that are mostly analog or SRAM and gain little from smaller transistors while inheriting worse wires. Splitting a design so that only the logic sits at the leading edge, and the rest sits on a mature node, is unglamorous and usually faster to tape out.
What to watch#
The honest metric for a node is not transistor density. It is the delay of a realistic critical path, including the wire, at the voltage the part will actually run. Foundries know this and their internal teams model it. Customers get the headline number and a design rule manual.
My view is that the industry should publish interconnect delay alongside device performance for each node, with the test structure described in enough detail to check. That would make node selection an engineering decision again rather than a procurement ritual. It would also force more honest conversations about when a design should stay put.
The transistor is not the villain here. It is doing its job. The wire is the part that has quietly stopped keeping up, and the sooner that becomes the center of the conversation, the fewer schedules will be lost to a timing report nobody expected.
More articles