AlphaChip

2024: Product Announcement

At the end of September 2024, Google announced AlphaChip, a specialized artificial intelligence system for chip design. The new neural network promises to significantly speed up the process of creating a crystal plan, as well as optimize chips in terms of performance, power and footprint.

It is noted that the crystal plan is traditionally the longest and most laborious phase of microchip development. For example, in the case of a GPU, this can take up to two years. The design of less complex products takes several months, but requires millions of dollars in costs, since a large number of highly qualified specialists are involved in the work. Google claims AlphaChip can solve this problem in just a few hours.

Freepik

AlphaChip uses reinforcement training based on simulating the trial-and-error learning process that people use to achieve their goals. AI performs actions in a given environment, studies results and learns from these experiences to improve the quality of work in the future. In the case of AlphaChip, the system treats the chip planning process as a kind of game, where one circuit component is placed on the grid in one move. The system improves as more layouts form, using a graph neural network to understand the relationships between components.

AlphaChip can design chips for various purposes. In particular, since 2020, AlphaChip has been used to develop Google's own AI accelerators - TPU products (tensor processors). In addition, this neural network was involved in the creation of MediaTek Dimensity 5G chips, which are widely used in smartphones.^[1][2]

Notes