Resistance In Advanced Packages Is Now A System-Level Problem

Key Takeaways

• Kelvin measurement, which has been in use for decades, is no longer sufficient for addressing resistance in complex chips.
• The problem is that resistance is no longer concentrated in transistors, and where it does show up isn’t always consistent or obvious.
• Traditional pass/fail approaches need to be replaced by more granular and flexible analytics and methodologies.

For most of its history, Kelvin measurement [1] solved a very specific problem. If you wanted to know the resistance of a device, you had to subtract the resistance of everything leading to it, such as leads, probes, cables, sockets. The solution was elegant and, for a long time, sufficient. Drive current through one path, sense voltage through another, separate stimulus from observation, and the device would reveal itself.

In today’s advanced packages, however, resistance no longer resides primarily inside transistors or neatly bounded test structures. It spreads across interfaces, between materials, and along paths that only exist temporarily during manufacturing and test. It drifts, accumulates and shows up differently, depending on where and when you look.

Engineers still measure resistance at final test, but the number frequently arrives too late to explain what actually happened upstream. Yield loss, performance drift, and reliability concerns often announce themselves quietly, long before devices fail outright or screening thresholds are crossed. At advanced nodes and in advanced packages, small resistance shifts are now routinely visible well before traditional pass/fail criteria are triggered.

“In advanced packages, the most common ‘escapees’ are marginal resistance-driven effects in the interfaces that are still within spec at time zero, but drift with thermo-mechanical cycling and real traffic,” said Nir Sever, senior director of business development at proteanTecs. “Microbump- and interposer-related degradation is a good example because you can pass a protocol or pattern test, yet still have a lane that is slowly losing integrity, which only becomes obvious when you track it continuously in mission mode and see the lane trend degrade.”

This changes the role of resistance measurement. It is no longer simply a value to be verified at the end of the line. It becomes a signal that must be preserved across time, across process steps, and across physical boundaries if it is to remain meaningful.

At its core, Kelvin measurement reduces this to a simple relationship:

That equation has not changed in a century. What matters is not the math itself, but the intent behind it. Drive current through a defined path and measure voltage somewhere else. Keep the act of stimulation separate from the act of observation so that parasitics, contacts, and the delivery path do not contaminate the result.

That logic still holds, but the world it was designed for no longer exists because modern devices don’t present themselves as tidy electrical endpoints. Classical Kelvin assumed stable contacts, well-defined current paths, and resistance concentrated in discrete locations. In contrast, today’s packages present distributed interfaces, transient conductive states, and resistance that shifts across temperature, time, and operating conditions. Kelvin measurements were designed for a world where the thing being measured could be cleanly separated from everything delivering the measurement, but that separation is no longer guaranteed.

What classical Kelvin assumes
Classical Kelvin measurement is often described as a wiring scheme, but that description misses the point. The four-wire configuration is merely an implementation detail. The deeper idea is philosophical. Kelvin measurement is a discipline built around the separation of a signal from the structures required to deliver it. Those kinds of separations used to be physical, but in advanced packaging, they are increasingly conceptual.

“Kelvin style methods stay valuable for controlled, absolute measurements at known nodes, especially early in development and for calibration,” said Sever. “What is changing is that many of the failure precursors we care about are no longer well represented by a static resistance number at a discrete test point.”

Classical Kelvin measurement rests on a handful of assumptions that are rarely stated because, for decades, they were almost always true. It assumes that the device under test is the dominant contributor to resistance. It assumes that contact resistance (CRES) is either negligible or stable enough to subtract out. It further assumes that the measurement interface itself does not meaningfully disturb the system being measured, and resistance can be localized in space and time to a single test event. Advanced packaging quietly breaks all four of those assumptions.

Contacts are no longer incidental. They are mechanical systems in their own right, subject to force, deformation, contamination, and wear. Interconnects span redistribution layers, substrates, and vertical transitions. Power and ground return paths run through parallel networks whose health cannot be inferred from a single node. Even the act of probing can change the very interface whose behavior is under scrutiny. As a result, what engineers often label as electrical noise is frequently something else entirely.

“When people say noise, usually they’re not talking about electrical noise,” said Jack Lewis, CTO at Modus Test. “What they’re really seeing is variation in the resistance of the interconnect itself. It’s not really noise. It’s the difference in the impedance or the resistance of the interconnect, and that’s where Kelvin comes in.”

As packaging scales outward, interface problems persist. They reappear in different geometries. When manufacturing shifts from wafer-based processes to panel-based substrates, the electrical questions follow the materials. Interfaces still charge, surfaces still accumulate potential, and measurements still depend on separating what is being observed from the structures delivering that observation.

“Right now, we’re moving to panel-level engineering,” said Lubek Jastrzebski, director of new product development at Onto Innovation. “All the problems experienced during IC building on wafers will be experienced during panel engineering, as well. Charging is a very important issue, and control of charge is a very important issue. This is what Kelvin is very good at.”

Once resistance becomes distributed, the Kelvin equation still holds. Resistance is the ratio of sensed voltage to forced current. What changes is the meaning of both terms. The forced current no longer travels through a single, well-defined structure. The sensed voltage no longer reflects a single interface. Both become aggregates of many contributors, some electrical, some mechanical, some historical. The math remains simple, but the system does not.

This is where misinterpretation begins. The measurement is often correct, but the framework used to interpret it is not. Engineers see variation, but the assumptions they are using to explain it no longer match the physical system generating the signal.

“Advanced packaging and multi-die designs are truly Z-dimensional. You’re no longer dealing with a single piece of silicon, but stacks of dies, interposers, substrates, and memory, all connected together,” said Sutirtha Kabir, executive director of R&D at Synopsys, in a recent conference presentation. “Power has to be pushed from the bottom of the stack to the top power-hungry blocks, and you have to account for voltage drop, thermal effects, and stress across the entire system. If you find those problems too late, you may not be able to fix them in silicon.”

Where resistance lives
To understand why resistance behaves differently today, it helps to look at where it actually lives. In advanced packages, resistance increasingly resides at interfaces between metal and dielectric, between the die and substrate, and between redistribution layers and vertical interconnects. These regions are electrically active, mechanically stressed, and often buried beneath other structures by the time a traditional electrical test occurs.

Some of these interfaces cannot be contacted at all without destroying them. Others change subtly as a result of thermal cycling, assembly, or repeated probing. The resistance they introduce is small but real, and it doesn’t belong to any single structure that can be isolated and measured cleanly.

“As pitches get finer, accessibility changes,” said Jeorge Hurtarte, senior director of product strategy for SoC marketing at Teradyne. “BGA balls at 300 to 400 microns are accessible. C4 and micro bumps at 50 to 80 microns are manageable. But at the hybrid bonding level, with direct silicon-to-silicon or copper-to-copper attachment, it becomes non-accessible. You have to have sacrificial paths or sacrificial pins just to test them.”

Seen this way, advanced packaging does not break Kelvin measurement. It exposes how dependent Kelvin measurement has always been a function of context. When resistance lived primarily inside devices, context could often be ignored. When resistance migrates into interfaces, materials, and interconnect stacks, context becomes inseparable from the measurement itself.

This is why non-contact Kelvin techniques have gained importance, not as replacements for electrical test, but as complementary views into behavior that would otherwise remain hidden until much later in the flow.

“Since it’s a non-contact measurement, it’s possible to look at individual processes much earlier in the fabrication flow,” said Dmitriy Marinskiy, applications manager at Onto Innovation. “Depending on the complexity, it can take many steps to get to the final device and test it there. The advantage of non-contact Kelvin probing is that you can monitor individual processes like CMP, wet chemical etch, ion implementation, and etching earlier in the production line.”

From a yield perspective, this shift has profound implications. Variability that once could be attributed to device performance increasingly originates elsewhere. As more functionality migrates into thinner films and engineered material stacks, electrical behavior becomes sensitive to properties that traditional test structures were never meant to observe directly.

“The technique allows us to characterize interfaces, where metal work function can change because of dipole layer formation between metal and high-k dielectric,” Marinskiy said. “Instead of using the metal to control the work function, people are now introducing the dipole in the dielectric itself. We can see this with the Kelvin probe because the dipole impacts the voltage, and therefore the signal.”

This is the quiet center of the problem. Resistance has become harder to interpret because its causes are distributed across a manufacturing system that no longer offers a single, stable vantage point. A resistance measurement taken at final test reflects not just the device itself, but everything the device experienced during assembly. That includes every mechanical stress applied during probing and every thermal excursion encountered in processing. The measurement may be accurate, but the cause is ambiguous.

“The problem with the interconnect is it changes from insertion to insertion,” said Lewis. “At any given insertion, a particular ground pin can go from normal to an outlier. And the magnitude of that change is much bigger than anything happening in your substrate or silicon trace. Suddenly, you can’t see the forest for the trees.”

Once resistance behaves this way, treating each measurement as an isolated event becomes risky. The same number can mean different things, depending on when it appears, where it appears, and what the structure carrying current has experienced since the last time it was observed. A 50-milliohm contact might be acceptable on one insertion and problematic on the next, not because the device changed, but because the mechanical presentation changed.

“Many of the failure precursors we care about are no longer well represented by a static resistance number at a discrete test point,” said Sever. “Embedded observability complements Kelvin by providing continuous, in-context monitoring of what really matters — functional margin under real stress.”

That realization sets the stage for a different way of thinking about Kelvin measurement. Its intent must be preserved even when the physical mechanisms that once enforced it have disappeared, rather than a wiring scheme or probe configuration.

Why hardware alone cannot solve this
Better probes can measure contact resistance more accurately, but they cannot explain why that resistance appeared or what it means in the context of everything that happened before.

“By the time you see the contact resistance issue, something’s already gone wrong and you’re already into the cost impact,” said Brent Bullock, test technology director at Advantest. “Right now, we don’t know exactly when probe burn events or socket burn events are happening. They’re very difficult to pinpoint, but the continuously running monitor should change the game. We should be able to pinpoint exactly when it happened and what the root cause of the burn event is.”

A voltage drop measured at wafer probe may correlate with yield loss at final test, but the mechanism connecting the two is not always obvious. Was it material quality? Process drift? Thermal excursion? Assembly stress? The test interconnect itself also becomes a confounding factor. Sockets and probe cards introduce their own resistance variability that can obscure the very signals engineers are trying to measure.

“Make sure your test interconnect is as good as it can be,” said Modus Test’s Lewis. “That’s a huge variable, and it injects noise into all the data that you’re looking at, and makes it harder to make decisions on what’s really going on.”

This is where hardware reaches its limit. The measurement noise floor is now often below the variability of interest. Engineers can see resistance shifts of a few milliohms, but those shifts exist within a background of mechanical variation, thermal drift, and contact inconsistency that is larger than the signal itself. More data does not automatically mean more truth. It can mean more noise if the framework for interpreting that data has not kept pace.

The problem is ambiguity, not sensitivity. Resistance measurements are now aggregates of many contributors — some electrical, some mechanical, some historical — and isolating the meaningful signal from the system noise requires more than better hardware. It requires correlation and analysis across time, process steps, and data domains.

Kelvin Everywhere
Kelvin Everywhere is the application of classical Kelvin measurement logic as a general engineering philosophy rather than a single test technique. It focuses on separating stimulus from observation across manufacturing, packaging, and system analysis to expose variability introduced by interfaces, materials, and process conditions. The approach enables earlier detection of degrading behavior in advanced packages, where electrical performance is tightly coupled to mechanical, thermal, and material effects.

Kelvin Everywhere is not about putting four-wire probes everywhere. It’s about preserving the separation of stimulus from observation, even when that separation can no longer be enforced by wiring alone. In advanced packaging, that separation must be preserved in the data, not just hardware.

The principle remains unchanged. Isolate the signal of interest from the structures used to deliver it. But when resistance is distributed across interfaces, materials, and process history, isolation becomes a question of correlation rather than configuration. Kelvin Everywhere means treating resistance as a signal that must be tracked across insertions, normalized against process context, and interpreted through the lens of what the device experienced, not just what it measured.

This shift demands statistical sensitivity to small deltas rather than reliance on absolute thresholds. Kelvin Everywhere acknowledges that context matters as much as the measurement itself.

It also demands design-aware interpretation. Resistance shifts in a power delivery network mean something different than resistance shifts in a signal path. A microbump under mechanical stress behaves differently than a solder ball under thermal cycling. The measurement techniques may be similar, but the interpretation must consider location, function, and history.

Kelvin measurement still depends on accurate hardware, clean contacts, and precise calibration, but it now also depends on the ability to preserve resistance sensitivity across manufacturing steps, across test insertions, and across the boundary between observation and inference.

Analytics and correlation
Once resistance becomes distributed, single-point measurements lose predictive power. What matters is how resistance behaves over time, across process steps, and in relation to other observables. This is where analytics and correlation become an essential extension of measurement.

“The early indicator is often a ‘degrading fault’ pattern, where eye margin erosion accelerates unexpectedly relative to nominal,” said Sever. “The safety and reliability framework we use is built around detecting degrading behavior before it crosses the minimum operating threshold, which is the difference between predicting risk and reacting to a failure.”

Wafer sort data correlates with final test. Inline metrology correlates with electrical signatures. Design intent acts as a filter, helping engineers distinguish between variation that reflects real margin loss and variation that reflects measurement artifacts. And resistance becomes a leading indicator when it is tracked continuously rather than sampled once.

“The practical answer is that you stop treating measurement as a single external observation and instead correlate multiple, independent internal signatures,” added Sever. “When you have on-chip Agents measuring timing or eye margins, voltage behavior, and workload context, you can determine whether a margin change aligns with a real physical phenomenon or whether it is a test-induced artifact.”

Drift detection depends on this kind of correlation. A gradual increase in contact resistance across multiple insertions suggests process degradation, and a sudden spike may suggest contamination or mechanical failure. The measurement might read the same in both cases, but the response should be different. Analytics allows engineers to separate signal from noise by comparing current behavior against a historical baseline, adjacent data, and design expectations.

Correlation is not causation, however, and this is where the risk lies. “There is a series of test challenges when we go from monolithic to advanced packaging, such as stability issues and debug-ability issues,” said Hurtarte. “The ASP of a GPU advanced heterogeneously integrated package is expected to reach over $25k by 2030, and AI accelerator companies cannot afford to throw away die or modules.”

This balance between power and caution defines the role of analytics in Kelvin Everywhere. Correlation helps engineers see patterns they would otherwise miss, but it does not replace the need for physical understanding, experimental validation, or engineering judgment. The goal is to preserve intent across a system that no longer offers a single, stable vantage point.

Challenges remain
Kelvin Everywhere increases sensitivity, but sensitivity without discipline creates noise. Several challenges remain unresolved, and acknowledging them is as important as describing the opportunity.

Calibration across domains is not standardized. Wafer probe, package test, and system-level insertion all measure resistance differently. Correlating those measurements requires either shared reference structures or statistical models that account for insertion-specific variability. Neither approach is universal, and both introduce error.

Data ownership and accessibility complicate correlation. Design, manufacturing, and test data often live in separate silos managed by different organizations. Connecting them requires infrastructure that many companies do not yet have, and sharing them requires trust that not all partnerships support.

“Machine learning shines in handling variability and uncertainty; things like lot-to-lot process drift or complex design-manufacturing interactions that physics-based models struggle to predict,” said Eduardo Castro, senior manager of R&D engineering at Synopsys. “However, explainability and governance will remain critical.”

Interpreting resistance shifts without over-flagging remains difficult. A resistance change that signals yield risk in one context may be benign in another. Setting adaptive thresholds requires either deep process knowledge or extensive historical data, and most organizations have one or the other, but not both.

“It’s hard to get that signal when you’re talking about something that’s very subtle and hidden within thousands or tens of thousands of needles,” said Bullock. “We’re just waiting on customers to get the volume data. But once we have that data, then we’ll be able to detect these things using regular analytics that will notice it just like any other drift.”

Organizational readiness is as much a barrier as technical capability. Kelvin Everywhere requires collaboration between design, test, and manufacturing teams that have historically operated independently. It requires investment in data infrastructure, analytical tools, and cross-functional workflows that do not produce immediate returns. The technical problems are solvable. The organizational problems are harder.

These are not reasons to abandon the Kelvin Everywhere approach. They are reminders that expanding sensitivity without expanding interpretive rigor risks trading one set of problems with another.

Conclusion
Kelvin measurement remains essential. Its intent — the separation of stimulus from observation — must survive even as the physical realities that once enforced it disappear. Engineers must think systemically about resistance, not as a discrete property of a device, but as a signal that accumulates across interfaces, materials, and time.

Kelvin Everywhere is an evolution, not a revolution. It extends a century-old principle into a manufacturing environment that was never designed to accommodate it. The challenges are real, but so is the necessity. Resistance no longer resides where it used to. Measurement and analysis must follow.

Reference

1. The Kelvin measurements referenced in this article pertain to the contact-based 4-wire Kelvin measurements, not the non-contact CPD/Kelvin-probe method.