Google DeepMind says its artificial intelligence is already helping design chips used in data centers and even smartphones. But some chip design experts are skeptical of the company’s claims that these AIs can plan new chip layouts better than humans.

He said the new method, dubbed AlphaChip, can design “superhuman chip layouts” in hours, rather than relying on weeks or months of human effort. anna goldie and Azaria Mirhoseiniaccording to researchers at Google DeepMind. blog post. This AI approach uses reinforcement learning to figure out relationships between chip components and receives rewards based on the quality of the final layout. However, independent researchers say the company has yet to prove that such AI can outperform expert human chip designers or commercial software tools, and they say they are unable to demonstrate that such AI can outperform expert human chip designers or commercial software tools, and that they believe that current state-of-the-art The company hopes to test AlphaChip’s performance on public benchmarks that include cutting-edge circuit designs.

“If Google provides experimental results for these designs, we’ll be able to make a fair comparison, and we hope everyone will accept the results,” he says. patrick madden At Binghamton University in New York. “Experiments take a day or two to run at most, and Google has nearly infinite resources. The fact that these results aren’t being provided speaks volumes to me.” He declined to comment.

Google DeepMind’s blog post says: update Google for 2021 nature A journal paper about the company’s AI process. Since then, Google DeepMind says AlphaChip has helped design three generations of Google’s Tensor Processing Units (TPUs). TPUs are specialized chips used to train and run generative AI models for services such as Google’s Gemini chatbot.

The company also claims that its AI-assisted chip designs outperform those designed by human experts and are steadily improving. AI accomplishes this by reducing the overall length of wire needed to connect chip components. This could reduce the chip’s power consumption and increase processing speed. Google DeepMind also said AlphaChip created the layout for a general-purpose chip used in Google’s data centers, while also helping MediaTek develop a chip used in Samsung’s phones.

However, the code published by Google lacks support for common industry chip data formats, which suggests the AI method is currently more suited to Google’s own chips, it said. . Igor Markovchip design researcher. “We have no idea what AlphaChip is today, what it does or doesn’t do,” he says. “We know that reinforcement learning requires two to three orders of magnitude more computational resources than techniques used in commercial tools, and typically lags behind. [in terms of] result. “

Markov and Madden criticized the original paper controversial Claim that AlphaChip outperforms anonymous human experts. “Comparisons to unnamed human designers are subjective, non-reproducible, and very easily fooled. Although it is possible that the human designer is not trying hard enough or is underqualified. , there are no scientific results here,” says Markov. “Imagine if AlphaGo were reported to have won against an unknown Go player.”

In 2023, independent experts who reviewed Google’s paper revoked his nature An explanatory article that initially praised Google’s efforts. The expert is andrew kern At the University of California, San Diego, Public benchmarking efforts When we tried to replicate Google’s AI methods, we found that they could not consistently outperform human experts or traditional computer algorithms. The best approach was commercial software for chip design from companies like Cadence and NVIDIA.

“Reinforcement learning appears to lag significantly behind the state-of-the-art in every benchmark that would be considered a fair comparison,” Madden says. “I don’t think that’s a promising research direction when it comes to circuit placement.”