Diffusion Network
Assembly Line
Intel GenAI For Yield
Diffusion networks are much better suited to the task. Real samples with added noise are used to train the model, which learns to denoise them. Crucially, diffusion networks in this application were able to replicate the long tails of the sample data distribution, thus providing accurate predictions of process yield.
In Intel’s research, SPICE parameters, used in the design phase as part of device simulation, are used as input for the deep learning model. Its output is the predicted electrical characteristics of the device as manufactured, or ETEST metrics. And the results show the model is capable of correctly predicting the distribution of ETEST metrics. Circuit yield is defined by the tails of this distribution. So, by correctly predicting the distribution of ETEST metrics, the model is correctly predicting yield.
The potential here is clear: better optimization of chip yields at the design stage means lower costs. Fewer mask respins, shorter development times, and ultimately higher yield would all be strong differentiators for foundries and design teams that can implement models into their PDK/design flows.