EST
Collaborate
research active

AI Chip Verification Suite

Building reusable verification environments and analytics to accelerate AI accelerator bring-up and coverage closure.

SystemVerilogPython automationMetrics pipelines

Context

As AI workloads demand specialized silicon, the verification of AI accelerator chips becomes critical to shipping reliable hardware. This project grew out of my daily work at Microsoft, where I identified opportunities to systematize verification approaches.

Problem

Verification teams often rebuild test environments from scratch for each new chip program, leading to duplicated effort, inconsistent coverage, and slow ramp-up times for new engineers.

Constraints

Must integrate with existing design workflows, support multiple verification methodologies (UVM, formal, constrained-random), and remain adaptable across different chip architectures.

Approach

Designed modular UVM environments with configurable stimulus generators, automated coverage collection pipelines, and health dashboards that surface debug insights in real time.

Impact

  • Reduced test debug time through targeted stimulus and health dashboards.
  • Improved confidence in floating point paths for production delivery.

Outcome

Reduced test debug time and improved floating-point pipeline confidence for production delivery. The framework is being adopted across verification teams.

Lessons Learned

Reusable verification infrastructure pays compound interest — the upfront cost of abstraction is small compared to the time saved across programs. Investing in observability (dashboards, metrics) changes how teams make decisions.

A focused initiative extending core verification work into reusable frameworks and knowledge transfer artifacts for future silicon programs.