Google DeepMind says its artificial intelligence has helped design chips that are already being used in smartphones as well as data centers. But some chip design experts are skeptical of the company’s claims that such AI could plan new chip layouts better than humans.
The newly named AlphaChip method can design “superhuman chip layouts” in a matter of hours, rather than relying on weeks or months of human effort, it said. Anna Goldie and Azalea MirhosseiniGoogle DeepMind researchers blog post. This AI approach uses reinforcement learning to identify relationships between chip components and rewards them based on the quality of the final layout. But independent researchers say the company has not yet proven that such AI can outperform professional human chip designers or commercial software tools. They are eager to see AlphaChip’s performance in public benchmarks involving current state-of-the-art circuit designs.
“If Google provided experimental results for these designs, we would be able to make fair comparisons and we expect everyone to accept the results,” he says. patrick madden from Binghamton University in New York. “It takes up to a day or two to run an experiment, and Google has almost infinite resources. The fact that these results weren’t available speaks volumes to me.” Google DeepMind declined to provide further comment.
Google DeepMind’s blog post states: update Google by 2021 nature Journal paper on AI processes in companies. Since then, Google DeepMind said AlphaChip helped design the third generation of Google’s Tensor Processing Unit (TPU), a specialized chip 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 constantly improving. AI achieves this by reducing the total length of wires needed to connect chip components. This is a factor that can lower chip power consumption and potentially improve processing speeds. And Google DeepMind says AlphaChip created the layout for general-purpose chips used in Google’s data centers and helped MediaTek companies develop chips used in Samsung phones.
However, Google’s publicly released code lacks support for common industry chip data formats. He said this suggests that AI methods are better suited to Google’s currently proprietary chips. Igor MarkovChip design researcher. “We don’t really know what AlphaChip is today, what it does and what it doesn’t do,” he says. “We know that reinforcement learning requires two to three times more computing resources than the methods used in commercial tools and typically lags behind. [in terms of] result.”
Markov and Madden criticized the content of the original paper. controversial They claim that AlphaChip outperforms an unnamed human expert. “Comparisons to unnamed human designers are subjective, unreproducible and very easy to game. Human designers may put in less effort or be underqualified. There are no scientific results here,” says Markov. “Imagine that AlphaGo reported that it had beaten some unnamed Go players.”
In 2023, an independent expert who reviewed Google’s paper will withdrawn his nature This is an op-ed article that originally praised Google’s work. The expert, andrew river At the University of California, San Diego Public benchmarking efforts Attempts to replicate Google’s AI methods have shown that they cannot consistently outperform human experts or existing computer algorithms. The best performing methods have been commercial software for chip design from companies such as Cadence and NVIDIA.
“In every benchmark that I think is a fair comparison, reinforcement learning seems to lag far behind the state of the art,” says Madden. “When it comes to circuit placement, I don’t think that’s a promising research direction.”
subject: