Skip to content
Theme
OnDevice Tools

Tech BusinessChips & Semiconductors

PVTW characterisation measures how far chips sit from their limits

Semiconductor Engineering argues that chip testing should record how much margin a device has, not just whether it passes. The PVTW approach adds workload to the usual process, voltage and temperature variables.

Illustration of a silicon wafer beside test equipment graphs showing voltage and temperature margins

Key takeaways

  • Semiconductor Engineering argues characterisation should measure how much margin a device has, not just whether it passes or fails.
  • PVTW extends process, voltage and temperature analysis with workload, revealing which software activity exposes a device's weakest condition.
  • Comparing known good and marginal device populations under identical conditions can hint at the physical mechanism behind failures.
  • Deep characterisation of long system-level tests can lead to much shorter production screens earlier in manufacturing.
In this article
  1. Why pass and fail is not enough
  2. Adding workload to the test variables
  3. Mapping the edges of the operating envelope
  4. Comparing good and marginal devices
  5. Turning long tests into short screens
  6. Margin as a continuous measurement

Why pass and fail is not enough

Chip testing in production is built around binary decisions, according to Semiconductor Engineering in a technical article published on 19 November 2026. A device either passes its specification or it does not, and it either ships or it does not. Characterisation, by contrast, asks a different question: not only whether a device meets its limits, but how much distance separates it from those limits.

The publication argues that this distinction grows more important as processors combine computing cores, memory, fast input and output links, several separate power domains and advanced packaging within one system. Five devices can all pass the same test yet sit at very different distances from failure. Under an identical workload, their minimum stable operating voltages may differ considerably, and production test would report all five as identical results.

Measuring that distance turns a narrow question, such as why a particular device failed, into a broader and more useful one about robustness across a whole population of parts.

Adding workload to the test variables

The article proposes PVTW as a framework for this work. The abbreviation stands for process, voltage, temperature and workload, extending the familiar PVT triplet used in chip characterisation with the software a device is actually running.

The suggested method is to find a device's voltage and frequency boundary with conventional automated test equipment, which is the machinery used to test chips in production. The same boundary is then measured again while the device runs a deterministic compute workload, meaning a repeatable program with known behaviour. The exercise is repeated with workloads focused on SRAM, memory, PCIe traffic and combinations of these.

If the boundaries differ between workloads, that difference points to which activity exposes the weakest condition in the design. It also suggests where deeper investigation should start.

Mapping the edges of the operating envelope

Illustration: privacy, temperature, market trend, news and the labels PVTW, Wafer Sort, Chips & Semiconductors

Rather than testing at a single nominal operating point, Semiconductor Engineering describes deliberately steering a device towards its limits. Voltage is swept at a fixed frequency, temperature and workload. Frequency is then swept while everything else stays constant. The sequence is repeated at different temperatures, and individual power rails are adjusted one at a time.

A compute-heavy workload, for example, can be characterised while the core power rail is varied and the memory and input and output rails stay fixed. A memory-heavy workload can be measured while a memory-related rail is swept. PCIe traffic can be kept running while electrical and thermal conditions change.

The result is a measurement of how much headroom a subsystem retains under a defined workload, rather than a simple statement of whether it works.

Comparing good and marginal devices

Population analysis adds another layer. The article suggests taking a group of known good devices and a group showing reduced margin in system-level test, a step that checks a chip inside a complete assembled system. Both groups are then characterised under identical PVTW conditions.

The comparison focuses on the shapes of the two groups' operating envelopes, which are the ranges of conditions each device can tolerate. Questions worth asking include whether the weaker group shows a consistent shift in voltage, whether the gap only appears at high temperature, or whether it emerges only when compute and memory activity happen at the same time. A particular power rail or frequency transition may also be where the two groups separate.

According to the article, the shape of the margin distribution can offer clues about the underlying physical mechanism causing the difference.

Turning long tests into short screens

One worked example involves a system workload that needs 45 minutes before marginal devices start producing errors. That is useful for learning about a design, but far too slow and costly to run as a production screen.

The proposed route is to use PVTW characterisation to understand why that long workload exposes the weakness. Once the mechanism is clearer, automated test equipment can hunt for a shorter electrical, structural or functional signature that correlates with the same marginal behaviour. A specific voltage and frequency corner might expose it within seconds, or a targeted functional pattern might reproduce the condition.

If the correlation proves strong enough, the shorter test could move into final test or even wafer sort, which happens earlier in manufacturing. The article sums up the learning flow as: system-level test discovers, PVTW characterises, automated test equipment correlates, and production test screens. The value of deeper system-level test, in this view, may lie less in running more of it and more in teaching engineers how to screen risk earlier and faster.

Margin as a continuous measurement

The article's author argues that margin data should be preserved rather than immediately reduced to pass and fail results. Minimum operating voltage, maximum stable frequency, thermal headroom, interface margin, power margin and workload margin can all be treated as continuous characterisation parameters.

Two devices may both satisfy today's production specification, yet one with 100 mV of margin and one with 10 mV carry very different implications for process variation, robustness and future design choices.

The argument does not stop at the silicon itself. The package around the chip has its own materials, manufacturing history and process variation. A follow-up article is planned on how 2D identification and unit-level traceability could link PVTW margin data back to the wafer and substrate panel positions a device came from.

Sources

  1. Beyond Pass/Fail: Characterizing Silicon Margin With PVTWSemiconductor Engineering

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.