AI Learns to Design Processors at the Level of Past Smartphones
At the WLA Forum, Chen Yunji discussed processors whose logic was created without manual coding. According to him, they have reached the level of mobile chips from the mid-2010s.
At the WLA Forum, Chen presented a system that independently creates the logical blocks of a processor and verifies their mathematical correctness. The testing environment identifies discrepancies and repeats the verification cycle hundreds of times per second.
The first experimental chip, Qimeng-1, was obtained by the team in 2021. In terms of processing speed and instruction density, it matched the Intel486 from the 1990s. By 2025, according to Chen, the successor Qimeng-2 will reach the level of central processors in smartphones from the mid-2010s.
The developers applied the same approach to machine learning: they studied over 100 common neural network models and identified recurring matrix and tensor operations. Based on these, an instruction set architecture was developed, which was licensed for early commercial computing acceleration chips.
Automation is also used in other semiconductor industry processes. Lam Research's neural network simulators, as claimed, accelerate etching calculations up to 30 times. Applied Materials analyzes over 1 million three-dimensional surface profiles of silicon wafers per hour. Digital models help optimize production without a large number of costly test runs.
The system does not yet replace leading engineers who create flagship chips. Currently, it is used for a closed-loop development and verification cycle of relatively simple computing cores, as well as for material analysis and equipment control.
