AI learns to design processors at old smartphone performance levels

Machine learning and optimization algorithms have long established themselves in EDA tools. Previously, they mainly assisted at the final stages, such as placing elements on a silicon substrate and routing conductors. Now, researchers are trying to delegate a more complex task to neural networks: the synthesis of the processor microarchitecture itself.

At the WLA international forum, research group co-founder Chen Yunji shared the results of developing processors whose internal logic was created without human-written source code. According to him, modern generative models have grown to the level of mobile processors from a decade ago. Designing chip logic requires high precision, as any error in the instruction pipeline turns silicon into an unusable piece of board.

• Neural networks have learned to create processor microarchitecture without writing source code.
• The experimental Qimeng-2 chip reached the performance level of mid-2010s smartphones.
• An automated testing environment checks logic blocks and resolves discrepancies hundreds of times per second.
• AI also accelerates material synthesis, etching simulation, and microchip defect analysis.

Closed-loop verification system

Attempting to assemble a processor solely through generative algorithms runs into a huge number of possible combinations of registers, buses, logic gates, and pipeline stages. Another problem is that the algorithm cannot simply invent an entire processor on the first try, and verifying such a scheme manually with standard tests is nearly impossible. Chen's team built a closed-loop system where the generator outputs logic blocks and an automated testing environment checks their mathematical correctness.

In 2021, using this methodology, they obtained the experimental Qimeng-1 chip, which matched the performance speed of a 1990s Intel486 processor. By 2025, its successor Qimeng-2 made a significant leap: the synthesized schemes reached the performance level of mid-2010s smartphone central processing units. The same analysis principles were applied to machine learning algorithms, studying over a hundred common neural network models.

Material synthesis and digital twins

In applied tasks, engineers increasingly use reverse engineering, where they first set target parameters such as mechanical strength, and the algorithm calculates the optimal chemical composition. At microchip manufacturing plants, processes like plasma etching require lengthy tuning, prompting manufacturers to introduce virtual modeling. Lam Research uses digital models with neural network simulators, speeding up etching process calculations by up to 30 times.

Applied Materials analyzes silicon wafer parameters during processing, evaluating over a million 3D surface profiles per hour. Equipment manufacturers link real working chambers with virtual copies to compensate for temperature and pressure fluctuations. Dedicated analytical agents take over defect analysis, reducing the time spent locating hidden micro-defects under an electron microscope from hundreds of hours to less than six.

Prospects of microelectronics automation

Automatic logic synthesis is not yet ready to completely replace leading architect engineers when creating flagship chips. Nevertheless, it already successfully closes the development and verification cycle for relatively simple computing cores. Combined with material synthesis and factory equipment control, it is becoming a full-fledged working tool for the semiconductor industry.

For the tech market and developers, the evolution of such systems introduces new optimization methods to the industry. While creating server-grade monsters entirely via AI is still distant, routine tasks such as synthesizing specialized blocks and analyzing manufacturing defects are being offloaded to algorithms, saving developers tremendous amounts of time and resources.