Nvidia’s leading scientist, William Dalley, reported that the corporation engages in a neural network at almost every step of the production of graphics accelerators. Previously, the adaptation of the catalog of typical cells to a fresh technological node required the efforts of eight developers over a period of about ten months. Now, the NB-Cell algorithm, built on reinforcement learning principles, performs an identical amount of calculations in one night using the power of a single graphics processor.

To support the engineering staff, branded language models Chip Nemo and Bug Nemo were created, incorporating all internal documentation and source code of previously released video chips. They play the role of curators, able to explain to newcomers the working principles of intricate nodes. During the setup phase, self-learning algorithms sometimes generate non-trivial architectural moves that outperform manual solutions by 20-30% in terms of area occupied, energy consumption, and speed.

However, the full autonomy of the speech is not yet discussed. Artificial intelligence helps to accelerate the procedure, but is not fully able to replace it.