One of the top minds in supercomputing has a warning for the West. Jack Dongarra, an industry luminary, Turing award laureate, and co-founder of TOP500, says the US is likely behind China in the exascale supercomputing race. It appears that a third Chinese exascale supercomputer, long thought to be halted indefinitely due to the impact of US sanctions, has come online. Meanwhile, the US currently only has two exascale machines, Frontier and Aurora, in operation.
When Dongarra says China might have the lead, it isn't taken lightly. China hasn't submitted its fastest exascale-class supercomputers to the Top500 list, the de-facto supercomputer listing and benchmarking process that's responsible for bringing clarity to how the world stands in processing capability, fearing that it will draw the attention of US regulators, thus triggering more sanctions. It follows that the official picture provided by the Top500 isn't an accurate representation of reality as it is missing data.
According to the TOP500, China comes in at a distant 7th (with the Sunway TaihuLight supercomputer) and 10th place (Tianhe-2) - trillions of calculations per second away from the top dog, the US-based (and AMD-powered) Frontier exascale computer. This isn't a picture of reality.
"It’s a well known situation that China has these computers, and they have been operating for a while," Dongarra told the South China Morning Post. "They have not run the benchmarks, but [the community has] a general idea of their architectures and capabilities based on research papers published to describe the science coming out of those machines."
The TOP500 highlights as of June 2023 show that China doesn't even crack the top 5 out of the world's supercomputers. The US, AMD-made Frontier takes top spot, while Japan's Arm-based Fugaku takes second. (Image credit: TOP500)
We know China has installed two exascale-class supercomputers — China doesn't submit its machines to the Top500 list, but it has submitted results for two of these machines for the Gordon Bell Prize. This yearly award recognizes "outstanding achievement in high-performance computing" based on the types of science run on the machines.
However, according to the SCMP article, it appears that a third machine, which hasn't been submitted for the prize, has now come also online.
China has submitted two machines for the Gordon Bell Prize: The Sunway OceanLight, which was developed by the National Supercomputing Centre in Wuxi, and the Tianhe-3 from the National Supercomputing Center of Tianjin, which was crunching AI workloads before it became mainstream.
(Image credit: HPC Wire)
However, there's also a third unnamed supercomputer, purportedly made by China-based Sugon, at the National Supercomputing Center in Shenzhen. This machine hasn't been submitted for the prize, and the project has long been thought to be halted indefinitely due to Sugon being placed on the blacklist in 2019. The company also lost access to the Hygon CPUs, a series of China-produced x86 processors based on AMD's Zen design, that it planned to use because US sanctions shut down AMD's joint venture that produced the processors. It's unclear which processors are being used for the machine.
The TOP500, being a voluntary listing, means that most won't step up to the task. And when we consider the current geopolitical climate, one can see it isn't conducive to transparency, openness, or a "head first" participation style.
"Maybe having the No.1 computer would make news and put China under the spotlight," Dongarra told the SCMP. "It can cause the US to take actions against China that would further restrict technologies from flowing into China."
It's interesting how Dongarra's observation maps onto reality: China having the top supercomputer would definitely make the news. And when it comes to tech restrictions being imposed or aggravated, we've had a couple of years of that already.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
“However, China is still the country which produces the most supercomputers. With domestic and Western-designed chips, supercomputers assembled in China are sold all over the world, including the US,” said Dongarra.
We've all felt the consequences of the sanctions at some point or another. Of course, there's still the open question of whether the sanctions are actually doing what they are intended to - restrict China's ability to catch up to the U.S. technologically and economically. Chinese companies being publicly bullish about recouping sanction-imposed losses in a year's time has to put at least a question mark over their effectiveness in the first place.
TOP500 Co-Founder Jack Dongarra recently commented on the sensitive geopolitical environment around the supercomputer listing and benchmarking organisation, drawing attention to potential missing data.
This great summary seems to have been lost in the article.
I think and worry more about intention than capability. No one is alarmed at the Finnish entry at the top of the chart.
I doubt that the unlisted computers were built to run climate models for future sustainability.
This great summary seems to have been lost in the article.
I think and worry more about intention than capability. No one is alarmed at the Finnish entry at the top of the chart.
I doubt that the unlisted computers were built to run climate models for future sustainability.
The problem with this article is If China truly had this third supercomputer capability, we would be seeing it all over the news as proof that China can overcome western sanctions and maintain competitiveness just like with the Huawei Mate 60.
The problem with this article is If China truly had this third supercomputer capability, we would be seeing it all over the news as proof that China can overcome western sanctions and maintain competitiveness just like with the Huawei Mate 60.
The article explains why that probably wouldn't happen.
It only works for certain kinds of tasks that involve very little communication between compute nodes, not a large amount of input data, and can be broken up into fairly small parcels of work. For everything else, you need a real supercomputer.
--
As for China's new ExaFLOPS-scale supercomputer, I think it's not too hard to compensate for being about 1 node behind, when you have essentially unlimited resources to throw at the problem. 2 nodes or more... that's where things will get interesting.
Either way, you can be pretty sure they're not going to be winning any efficiency records on these things.
As for China's new ExaFLOPS-scale supercomputer, I think it's not too hard to compensate for being about 1 node behind, when you have essentially unlimited resources to throw at the problem. 2 nodes or more... that's where things will get interesting.
Either way, you can be pretty sure they're not going to be winning any efficiency records on these things.
America put people on the moon, built an enormous nuclear-tipped ballistic missile force and created the internet with computers less powerful than the ones a Foxconn janitor has in his pocket. I don't think being limited to a 7nm process will hold them back too much.
China is permitting two new coal-fired power plants a week and 22 nuclear power plants under construction. I don't think the energy efficiency angle will give them much pause.
America put people on the moon, built an enormous nuclear-tipped ballistic missile force and created the internet with computers less powerful than the ones a Foxconn janitor has in his pocket.
That has nothing to do with anything. A lot of the problems we haven't solved require vast amounts of compute power, which usually has something to do with why we haven't solved them.
Co BIY said:
China is permitting two new coal-fired power plants a week and 22 nuclear power plants under construction. I don't think the energy efficiency angle will give them much pause.
They're adding so much energy-generation, in large part, because they have a large population. The central government actually tried to stop the build-out of coal-fired power generation, a few years ago, but lost a legal battle and had to relent.
Oh, and another thing that happened, recently, is the Yangtze River dried up in a historic drought, which was a major source of power generation:
https://www.theguardian.com/world/2022/aug/22/china-drought-causes-yangtze-river-to-dry-up-sparking-shortage-of-hydropowerI'm sure some of the new capacity is to backstop a recurrence of that. During a heat wave is usually when energy consumption is peaking, too. It was so bad that they had to temporarily shut down factories in the region.
Anyway, I basically agree that they can & will build whatever amount of power generation capacity needed to run their supercomputers and datacenters, which is part of what worries me. However, I do think you face some scaling challenges as supercomputers grow physically larger. So, it's not a simple "numbers game" where you can necessarily make up in quantity what your nodes lack in width and speed.
