Why the move to advanced packaging is reshaping how the industry collects, moves, and acts on test data, and how Data Feed Forward turns upstream measurements into downstream intelligence.
For most of the industry’s history, the lever for semiconductor performance gains was process-node scaling. That is no longer the whole story. As one recent industry analysis put it, advanced packaging has now displaced node scaling as the primary lever for performance gains. The consequence reaches well beyond design and assembly; it lands squarely on test.
In a monolithic world, test optimization was hard but bounded: a single integrated die was validated, and the data and decision boundaries stayed relatively contained. Chiplet-based architectures break that containment. They multiply the number of components in a package, the number of transitions between process steps, and the number of opportunities to collect and act on test information. The question is no longer how much test data a flow can generate. It is whether that data can be moved to the right place, at the right time, across a distributed supply chain, and applied to a live production decision.
In this context, PDF Solutions proposes the solution concept it calls Data Feed Forward (DFF): a framework for collecting, transforming, transporting, and applying data generated early in the manufacturing and test process so it can drive smarter decisions downstream.
A chiplet-based device is not validated once. It is graded across multiple dies, multiple stages of assembly and integration, and multiple test insertions. A typical flow now spans fab operations and process control monitoring (PCM), wafer sort, assembly and bonding, package test, final test, and burn-in or system-level test (SLT). Each insertion produces a useful signal, but each is also a progression from what happened upstream.
Two structural problems follow. First, more dies and more interfaces mean more failure modes and more coverage requirements; each chiplet brings its own test needs. Second, the decisions become interdependent: a result at wafer sort can directly influence what should happen at package test, final test, or SLT, and errors left uncorrected compound across the full flow.
The conventional, step-by-step model handles this poorly because each insertion tends to operate on its own local data. If something important is learned at wafer sort, there is rarely a mechanism to make that knowledge available later at final test or system-level screening. The challenge, then, is not simply that there is more data; it is that effective downstream decisions could increasingly leverage upstream context and usually don’t.
The deeper barrier is structural. Semiconductor manufacturing and test are highly distributed: stages occur in different factories, different countries, and across different organizations: design-house, wafer fab, OSAT, final OEM, and the field. Each site runs its own data stack.
In principle, most teams understand that upstream data has value. In practice, it is fragmented, delayed, or simply inaccessible at the moment a downstream decision must be made. When that happens, engineers fall back to a simpler, more isolated decision model, and three things degrade at once: visibility suffers because downstream teams lack the full device history; traceability weakens because final outcomes are harder to connect to earlier test cycles; and AI/ML deployment is constrained because models can only see the data available at the immediate insertion where they run.
So before adaptive test, optimization, or advanced modeling can deliver, there is a more basic operational question to answer: getting the right data to the right place, at the right time, across a distributed supply chain.
Data Feed Forward (DFF) is the answer to that question; it is an operational concept, not a storage concept. The point is not to preserve data for later analytics. The point is to make upstream data useful in live downstream decisions.
In practice, results from an early insertion, say wafer sort, are converted into something a later step can act on: engineered features, inferred device attributes, model predictions, quality indicators, or routing recommendations. Those outputs are then delivered to downstream processes, which may sit at remote test facilities or external supply-chain nodes. The effect is to turn upstream test results into downstream process intelligence, creating continuity across stages that would otherwise be disconnected.
That continuity is the precondition for everything else. Once it exists, a far more intelligent and adaptive test methodology becomes possible.
Expressed as architecture, a best-practice DFF implementation runs as five operational layers:
Collect. Ingest data from wafer sort, probe, and early test insertions; parametric data, inspection results, binning data, waveform signatures, and process context.
Transform. Convert raw data into something actionable through feature engineering, rule generation, model inference, or decision thresholds. Raw data is usually too heavy and too unstructured to use directly downstream, which makes this layer essential rather than optional.
Transport. Move the transformed outputs securely across sites and partners. In distributed environments (remote facilities, external partners), secure, reliable delivery is a hard requirement, not a convenience.
Apply. Feed the outputs into downstream operations to drive concrete actions: smarter screening, adaptive limits, trim-target prediction, routing, or test-suite selection. In practice this resolves decisions like bin, skip, adjust, or flag.
Write back. Record outcomes so the system stays traceable and continuously improving. This closes the loop: predictions and actual results can be compared and used to refine algorithms and retrain models over time.
That closed-loop structure is what makes DFF scalable and operationally meaningful. Intelligence is not pushed forward once, the system learns and adapts.
A framework needs production-grade plumbing for both data movement and model deployment. PDF Solutions positions two products for this. Exensio Test Operations handles secure data collection, data management, process monitoring, quality control, edge deployment, and supply-chain connectivity. Because it typically interfaces directly with test equipment, it captures the highest-fidelity, most timely source of test data and can apply rules to watch the process for excursions that would otherwise compromise data quality.
Exensio StudioAI extends that with model-ops capability across the full machine-learning lifecycle: building, training, deploying, and governing models, with the option to use open-source algorithms or “bring your own.” Because it can draw on aligned, trusted manufacturing data already present in the Exensio system, data scientists can focus on modeling rather than on gathering data or engineering a deployment path. Together, the two turn DFF from an idea into operational deployment at the edge.
The payoff can be grouped into three categories:
Efficiency. Much downstream testing is designed conservatively because the downstream step does not know what was already learned upstream, so one size has to fit all. Bring that knowledge forward reliably and coverage can be right sized per device. A frequently cited example is selective burn-in avoidance: where predictive confidence is high and the process well controlled, predicted-good devices can skip unnecessary stress steps, cutting time and cost while preserving product objectives, and freeing burn-in capacity for devices genuinely at risk of early failure. A no-risk variant is narrowing a characteristic search, such as Vmin, to a boundary already identified upstream, reducing test time without added risk. Efficiency here is less about blunt cost reduction than about running the right test on the right device at the right time.
Quality. This may matter even more than efficiency. Predicting trim targets before a step executes reduces mis-trims, a clean case of upstream information directly improving a downstream operation. Monitoring characteristic signals across insertions and sites surfaces drift early, which can flag test instability or, more seriously, device changes from premature aging or stress that point to a long-term reliability risk. Aligning thresholds across facilities and partners yields more consistent outcomes, and better upstream screening reduces escapes. The broader point: overall quality is a supply-chain property, not a single test outcome, and DFF is what makes it addressable from end to end, shifting quality management from reactive to predictive.
Performance. This is where DFF connects most directly to AI-driven methods, and where “performance” means higher-performing decisions, not just faster devices. The most advanced inline models are often limited because they see only the data at their immediate step. Enrich them with upstream or historical context, combining wafer-sort signatures with package-level observations, or blending process history with test measurements, and predictions become more accurate and more context-aware. That supports smarter grading and routing for expensive advanced packages that, given their cost, cannot be scrapped but must be characterized and graded to a suitable application. System-level test benefits too: fed-forward data can dynamically focus the SLT suite on specific device functionality. And the same logic extends to assembly optimization, matching the performance of sub-components to their target application to maximize the price-performance of the finished device.
The throughline across all three is the same. Chiplet-based packaging requires connected data across the supply chain. DFF is the operational backbone that makes upstream data usable downstream, and its benefits span efficiency, quality, and performance simultaneously. In advanced packaging, competitive advantage increasingly depends not on how much data a flow can collect, but on how effectively intelligence is moved forward and applied where it matters.
Put plainly, this is a shift from data accumulation to data activation, in line with production. In the advanced-packaging era, the advantage goes to the teams that feed intelligence forward, not just the ones that collect it.
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Leave a Reply