Congratulations to Asma Mohsin and Muhammad Ali Farooq at Rapid Silicon for winning the third-place prize in the Efabless AI-Generated Design Contest!

Welcome to the Q&A with Asma Mohsin, the third-place winner of the Efabless AI-Generated Design Contest. Asma will be discussing her work on Model Predictive Control (MPC). The design is used to predict future behavior and optimize control actions for a regulator control circuit provided in MATLAB code to ChatGPT-4 and then implemented with prompts in Verilog.

How did you implement your project? What challenges did you run into?

To develop the MPC, the coefficient matrix, which captures the dynamics of the converter, was calculated using MATLAB. Additionally, a reference model was constructed in Simulink. This reference model provided a benchmark for the desired behavior of the controller, allowing for performance evaluation and comparison against the actual system response. In order to implement the MPC in hardware, the coefficients obtained from MATLAB were converted into fixed-point numbers. We found this to be impossible to implement via AI (both Bard and ChatGPT). To accomplish this, an online tool developed by GitHub user Chummersone was employed. This conversion is necessary to ensure compatibility with the target hardware and to facilitate subsequent implementation steps.

Next, the MATLAB code for the MPC was given as input to Google Bard, to develop an error-free module capable of carrying out the required control function. This was an iterative process involving refining the code, incorporating feedback from Google Bard, and improving the module until the desired performance was achieved. We faced some issues on this stage, since sometimes Bard would deliver nonsensical implementations which had to be carefully scrutinized.

Once the module was developed, it was tested using handwritten testbenches to verify its correctness and functionality. These testbenches involve designing specific input scenarios and comparing the module's output against expected results. Finally, the Verilog was converted to GDSII format using OpenLane.

What did you learn from this experience?

We honestly were not expecting a win on the module we submitted. We were more interested in gauging the capabilities of generative AI for developing application specific products. We realized that while generative AI might not be able to perform the full end to end flow yet, it is still an absolute game-changer for development and debugging.

How would you extend your project or what would you do next time?

In the future, we hope to implement a larger horizon MPC for longer prediction windows, for different applications. Instead of baking the coefficients into the design, we plan to implement a re-configurable module that can adapt its functionality based on the operating conditions.