![\"\"](\"https:\/\/images.newscientist.com\/wp-content\/uploads\/2024\/10\/02195809\/SEI_224192044.jpg\")
<\/div>Can AI design a chip that\u2019s more efficient than human-made ones?<\/p>\n
Yuichiro Chino\/Getty Images<\/p>\n<\/div>\n<\/figcaption><\/figure>\n<\/p>\n
Google DeepMind says its artificial intelligence has helped design chips that are already being used in data centres and even smartphones. But some chip design experts are sceptical of the company\u2019s claims that such AI can plan new chip layouts better than humans can.<\/p>\n
The newly named AlphaChip method can design \u201csuperhuman chip layouts\u201d in hours, rather than relying on weeks or months of human effort, said Anna Goldie<\/a> and Azalia Mirhoseini<\/a>, researchers at Google DeepMind, in a blog post<\/a>. This AI approach uses reinforcement learning to figure out the relationships among chip components and gets rewarded based on the final layout quality. But independent researchers say the company has not yet proven such AI can outperform expert human chip designers or commercial software tools \u2013 and they want to see AlphaChip\u2019s performance on public benchmarks involving current, state-of-the-art circuit designs.<\/p>\n \u201cIf Google would provide experimental results for these designs, we could have fair comparisons, and I expect that everyone would accept the results,\u201d says Patrick Madden<\/a> at Binghamton University in New York. \u201cThe experiments would take at most a day or two to run, and Google has near-infinite resources \u2013 that these results have not been offered speaks volumes to me.\u201d Google DeepMind declined to offer additional comment.<\/p>\n <\/p>\n Google DeepMind\u2019s blog post accompanies an update<\/a> to Google\u2019s 2021 Nature<\/em> journal paper<\/a> about the company\u2019s AI process. Since that time, Google DeepMind says that AlphaChip has helped design three generations of Google\u2019s Tensor Processing Units (TPU) \u2013 specialised chips used to train and run generative AI models for services such as Google\u2019s Gemini chatbot<\/a>.<\/p>\n The company also claims that the AI-assisted chip designs perform better than those designed by human experts and have been improving steadily. The AI achieves this by reducing the total length of wires required to connect chip components \u2013 a factor that can lower chip power consumption and potentially improve processing speed. And Google DeepMind says that AlphaChip has created layouts for general-purpose chips used in Google\u2019s data centres, along with helping the company MediaTek develop a chip used in Samsung mobile phones.<\/p>\n But the code publicly released by Google lacks support for common industry chip data formats, which suggests the AI method is currently more suited for Google\u2019s proprietary chips, says Igor Markov<\/a>, a chip design researcher. \u201cWe really don\u2019t know what AlphaChip is today, what it does and what it doesn\u2019t do,\u201d he says. \u201cWe do know that reinforcement learning takes two to three orders of magnitude greater compute resources than methods used in commercial tools and is usually behind [in terms of] results.\u201d<\/p>\n Markov and Madden critiqued the original paper\u2019s controversial<\/a> claims about AlphaChip outperforming unnamed human experts. \u201cComparisons to unnamed human designers are subjective, not reproducible, and very easy to game. The human designers may be applying low effort or be poorly qualified \u2013 there is no scientific result here,\u201d says Markov. \u201cImagine if AlphaGo reported wins over unnamed Go players.\u201d<\/p>\n In 2023, an independent expert who had reviewed Google\u2019s paper retracted<\/a> his Nature<\/em> commentary article that had originally praised Google\u2019s work. That expert, Andrew Kahng<\/a> at the University of California, San Diego, also ran a public benchmarking effort<\/a> that tried to replicate Google\u2019s AI method and found it did not consistently outperform a human expert or conventional computer algorithms. The best-performing methods were commercial software for chip design from companies such as Cadence and NVIDIA<\/a>.<\/p>\n \u201cOn every benchmark where there\u2019s what I would consider a fair comparison, it seems like reinforcement learning lags behind the state of the art by a wide margin,\u201d says Madden. \u201cFor circuit placement, I don\u2019t believe that it\u2019s a promising research direction.\u201d<\/p>\n Topics:<\/p>\n<\/section><\/div>\n