Navigating the complexities of maximizing efficiency in random testing for designs with multiple operational modes is a formidable challenge. Achieving comprehensive coverage across such varied designs necessitates running multiple randomized regression tests for each mode, consuming substantial verification and compute resources and time for both regression runs and subsequent result analysis. An opportunity exists for the integration of machine learning into this process.
Verisium SimAI revolutionizes regression testing by analyzing an initial batch of tests, including the number of tests and seed variations per test, to identify the coverage bins activated during these tests. Given the randomized nature of these tests, some coverage bins may be hit repeatedly while others remain untouched. SimAI utilizes this data to first curate and then create an efficient suite of tests to maximize coverage bin hits and minimize redundancies, allowing for a more resource-efficient "curated" random regression process.
SimAI advances beyond traditional coverage grading methodologies. Traditional grading often evaluates the contribution of individual runs to overall coverage without considering the interplay between specific variables and the test environment. This approach is not only broad but lacks adaptability to changes in test conditions. Verisium SimAI, conversely, employs machine learning models to dissect how each random variable influences overall coverage, enabling the creation of new tests tailored to explore specific variables and their combinations more effectively. This granularity enhances test efficiency and yields richer analytics for design verification (DV) engineers.
The Cadence Silicon Solutions Group (SSG)applied this innovative approach by setting up an AI-driven workflow based on an initial regression run comprised of a single Datapath over 2100 runs. After processing this initial data set through the machine learning model, an optimized verification session input format (VSIF) command was generated. This optimized VSIF demonstrated remarkable efficiency when rerun and analyzed, reducing the number of regression tests to a third of their original count without compromising coverage quality. This optimized approach achieved broader coverage with fewer tests and highlighted areas where coverage is inherently challenging due to design complexities.
The chart presented below illustrates the performance of the SimAI regression model. The Base and optimized (opt) regression models peak quickly. Initially, Base1 addresses 422 holes, with Base2 contributing an extra 69 holes and Base3 covering 35 holes. In total, 536 holes are accounted for by ML techniques, compared to 526 holes addressed by the Base method. This results in a gap of 66 holes, a difference that is challenging to bridge through further Base regression efforts.
While some hard-to-reach bins remain uncovered—highlighting limitations in addressing the intricacies of certain design elements, such as IDE blocks in non-IDE regressions—the results are promising. The application of Verisium SimAI has significantly shortened the verification cycle by enhancing test efficiency and coverage, underlining the potential for continued advancements in machine-learning-assisted verification processes.
The success of Verisium SimAI in streamlining design verification through smart, data-driven strategies showcases the potential of machine learning in overcoming traditional testing challenges. As the industry evolves, leveraging such innovative technologies will enhance efficiency and achieve faster, more accurate design verification processes. Cadence is committed to pioneering the integration of machine learning into verification flows, ensuring that engineers can unlock peak throughput with minimal bugs. Learn more about Verisium SimAI.