|Conference:||Verification Futures 2023 (click here to see full programme)|
|Presentation Title:||Closing Functional Coverage on A Compression Encoder with Deep Reinforcement Learning|
The simulation based constrained random coverage driven functional verification is one of the most prevalent paradigms used on RTL designs today. While greatly enhancing the verification process, reaching full functional coverage, a necessary step for verification closure, remains a bottleneck. This is mainly due to the necessary manual intervention in the process, consisting of the verification team adjusting stimulus randomization constraints and developing directed tests. This paper tackles this issue by exploring the use and integration of a deep reinforcement learning engine in such a verification environment
Specifically, the technique demonstrated here is a deep learning approach to a Q-Learning variant named Deep Q-Networks, a similar approach to the one developed in the work of DeepMind Technologies to play Atari games. The method is adapted to fit seamlessly into a constrained random coverage driven functional verification environment. The solution development stages along with the simulation results are detailed in this paper. We conclude by making suggestions on which designs may be good candidates for this novel technique
I am an accomplished design and verification ASIC engineer. I have worked for companies like Intel, ARM and Imagination Technologies in leading roles. I recently completed a PhD researching the use of AI algorithms for RTL verification.