I follow the (sometimes) weekly translations of Chinese-language musings on AI and related topics sent by Jeff Ding, a Ph.D. candidate in International Relations at the University of Oxford and researcher at the Center for the Governance of AI at Oxford’s Future of Humanity Institute. Those interested in AI research and its development in China should subscribe to this newsletter.
This week was interesting because it identifies the rollout of AI computer centers across China and indicates concern that the centers may not have sufficient demand:
“The goal: Let the computing power flow like tap water (让算力像自来水一样流淌)
• AI computing centers as “essential infrastructure” in all parts of the country. As the article reports, Xi’an, Xuchang, Nanjing, Hangzhou, Guangzhou, Dalian, Qingdao, Changsha, Taiyuan, Nanning, are among the cities that have started building or are planning to build computing centers to support AI applications.
• Four such computing centers have already been built. I think the PCL supercomputing center in Shenzhen (ChinAI #73) is one of them? ***Bonus points to the ChinAI reader that can track down the others.
The problems are twofold:
• 1) Price chaos — In one city, the construction cost for a computing center with performance of 100 PFlops (100P) at 16-bit precision is 75 million RMB. In another city, a computing center with the same specifications costs 450 million RMB, a difference of 6.2 times.
• 2) Confusion over how to benchmark compute clusters — different applications have varying requirements for precision. For instance, AI model training mainly uses 32-bit single-precision; AI inference (model implementation) can use 16-bit or lower. By contrast, some scientific calculations, such as weather forecasting or drug discovery, require higher 64-bit double precision. In the current rush to build computing centers, there’s been confusion over these different precision requirements. Specifically, the piece calls out the inflated prices for computing centers with high peak performance metrics (measured in PFlops) but low precision: these are deceptive gimmicks that “pass off fish eyes as pearls” [鱼目混珠] and can’t meet industrial needs.
• The report warns, “If these two problems are not resolved, the smart computing centers built will not match the true value in price, nor can it meet the corresponding demand, which will inevitably cause waste of resources and hinder the development of the industry.”
What’s the potential solution?
• The report emphasizes standardization and stable benchmarks, specifically highlighting efforts by the Chinese Academy of Sciences AI Industry-University-Research Innovation Alliance [中科院人工智能产学研创新联盟]. At the World AI Conference 2021, this CAS alliance released a new generation AI computing platform, which aimed to set the standard for intelligent computing centers.
• The key here is that many AI application scenarios, including material design and drug discovery, require a combination of AI and high-precision scientific computing. Toward that end, this platform “supports a multi-chip combination of CPUs, general-purpose GPUs, and dedicated AI acceleration chips, providing computing power covering various precisions, and can be competent for simulation, training, inference, and other AI full-chain application requirements.
• As for stabilizing prices, the CAS alliance gave out this guidance: “After integrating a series of factors such as storage, energy consumption, development, customization, and data scheduling, as well as plugging in clear algorithm standards, for an intelligent computing center with 5P double-precision computing power (64-bit), 25P single-precision computing power (32-bit), and 100P half-precision computing power (16 bits), the resulting infrastructure price is about 100 million-150 million RMB.”
Dig Deeper
Okay, I know we’re already in the weeds but let’s drill down even more and add some historical context. I think we can uncover a similar theme — impressive top-line numbers paired with underutilization — in China’s previous efforts to build supercomputers.
• See this 2010 Science article on Dawning 5000A, which was once China’s fastest supercomputer: “Only 1% of the applications on China’s previous speed champ, the Dawning 5000A at the Shanghai Supercomputer Center, use more than 160 of the machine’s 30,720 cores. For comparison, 18% of the applications running on Oak Ridge’s Jaguar XT5 use 45,000 to 90,000 of the machine’s 150,162 cores, according to a presentation at last year’s announcement of China’s top 100 fastest computers. ‘A supercomputer without software is like a wild horse without a harness,’ says Zhang Yunquan, a parallel computing researcher at the Institute of Software of the Chinese Academy of Sciences in Beijing. ‘Its horsepower is wasted.’”
• Brian Tsay, in a 2013 SITC Bulletin piece, writes, “it is not easy to write code that can actually utilize all the computing power that an HPC (high-performance computing) system has to offer. The result is that supercomputers can be left idle for long periods of time, raising the question of whether China even needs greater computing capacity.”
• My favorite passage from Brian’s piece, which discusses China’s Tianhe-2 (TH-2), once the fastest supercomputer in the world: “For example, in response to the notion that the TH-2 will be used to improve China’s automobile industry, a professor at Tsinghua University’s department of automobile engineering commented, ‘I have never heard of Toyota or Daimler or any major carmaker using a supercomputer to design their cars […] It is like running after a chicken with an axe. It is quite unnecessary.’”
Overview by Tim Sloane, VP, Payments Innovation at Mercator Advisory Group
