Introduction
As Very Large Scale Integration (VLSI) technology advanced in the late 1980s, the disparity between internal transistor counts and external I/O pins created a "visibility crisis." Internal components became increasingly difficult to control or observe, complicating manufacturing verification. Traditional Design for Testability (DFT) solutions, such as Level Sensitive Scan Design (LSSD), were often "heavy-handed," imposing rigid rules that added significant hardware overhead and performance penalties, even in areas that were already inherently testable.
This paper introduces an AI-driven "DFT advisor" that shifts away from one-size-fits-all rules. Instead, it employs a surgical approach, modifying a circuit only when a test generator specifically fails to verify a component. By using these failures as a filter, the system minimizes the area and performance overhead typically associated with making chips testable.
Methodology: The Failure-Driven Loop
The proposed system operates through a four-step, iterative loop designed to provide the minimum necessary modifications:
- Identify: An external test generator identifies "untestable" components.
- Analyze: The system attempts to generate a test for these specific components and analyzes the root cause of the failure (e.g., an inability to provide a specific sequence to a ROM input).
- Select: Based on the failure's nature, the system selects a repair strategy from a library of DFT techniques.
- Modify: The circuit is updated, and the process repeats until full testability is achieved.
Technical Innovations
A core innovation is the use of Compound Component Templates. The system recognizes that existing, disparate components can be combined to fulfill a testing requirement. For instance, it can identify that a register and an incrementor already present in the design can be looped to function as a counter for testing other logic.
Furthermore, the system prioritizes Additive Modifications—adding functionality (like a "shift" mode) to existing registers—rather than replacing them entirely. It also utilizes Constraint Relaxation in a hierarchical search, starting with strict requirements for low-overhead designs and only relaxing them if no solution is found.
Key Results and Case Studies
The system was evaluated on benchmark circuits, most notably the MAC-2 textbook microprocessor. The MAC-2 presented a challenge due to its complex sequencers and datapaths.
In the MAC-2 example, the system correctly identified that the sequencer components were inaccessible. Rather than applying a blanket scan-design approach, it suggested highly focused changes:
- Connecting the existing Multiplexor, Incrementor, and uPC into a loop to provide test patterns to the ROM.
- Adding a "shift" function to the microinstruction register (uIR) to allow the ROM's output to be observed externally.
This surgical approach resulted in significantly lower hardware overhead compared to traditional rule-based systems, as it repurposed existing "live" silicon rather than adding "dead" silicon dedicated solely to testing.
Insights and Limitations
The primary insight of this work is that testability is relative to the intelligence of the test generation algorithm. If a generator can handle a complex structure, no modification is needed. By treating failure as a filter, the system ensures that every added gate or wire serves a specific, proven need.
However, several limitations remain. The search process is computationally intensive, potentially requiring more hierarchical representations for modern, large-scale chips. Additionally, the system lacks "layout knowledge," meaning it cannot account for the physical routing or area impact of its suggested modifications. Finally, because it is failure-driven at the component level, it is not designed to suggest high-level architectural changes, such as restructuring a global bus.
Despite these constraints, the research demonstrates a powerful shift toward intelligent, minimal-overhead design for testability, prioritizing the repurposing of existing hardware over the addition of redundant structures.
Summary updated