China Supercomputer runs fast and performs well, but there is still a gap between software and ecology.

The "Sunway TaihuLight" supercomputer. Visual China Information

Liang He

Recently, China's self-developed supercomputer "Sunway TaihuLight" once again clinched the top spot at the 2017 International Conference for High Performance Computing, marking the third consecutive championship for a domestically developed Chinese supercomputer system in the world's top supercomputing rankings. Coupled with the six-time wins of the Tianhe-2 before it, China's supercomputers have truly shone on the global stage.

Despite its impressive "speed" and performance, however, the spotlighted supercomputer has a significant短板in its practical applications.

"Sailing out to sea" is not always a winning strategy.

Before the birth of "Sunway TaihuLight," China's Tianhe-2 was once the "dominant force" in high-performance computing worldwide. It claimed the top spot on the global Top 500 list six times and found extensive use in research fields such as meteorology, climatology, biomedicine, and新材料development. However, it also faced criticism. The processors in the Tianhe series relied on foreign technology: Tianhe-1 used Intel CPUs paired with Nvidia GPUs, while Tianhe-2 employed Intel GPUs combined with Intel many-core processors.

"The mature chip technology and rapid progress of its performance have been seen as a development strategy, but the realization of autonomy and controllability has not been achieved," Dr. Lin Gan from Tsinghua University's Department of Computer Science told a reporter from Science and Technology Daily on July 28th. A series of measures announced by the U.S. regarding embargoes on China's core processors also directly obstructed the advancement of the Tianhe system.

"Through 'borrowing' from abroad, Tianhe-2 surpassed U.S. systems like Titan, but in critical sectors such as national defense security, the country demands self-reliance," Dr. Gan noted. Key technologies dependent on others, even if mastered over the long term, significantly reduce their impact at the application level.

China's Core Enables a Jedi "Counterattack"

Why did the "Sunway TaihuLight" reaching the world provoke such excitement for China's supercomputing sector? The reason lies in its adoption of China's自主研发core, the "Sunway 26010" heterogeneous heterogeneous processor, achieving a "Jedi-style counterattack." To the public, this translates into a complete breakthrough in terms of自主controllability, peak speed, continuous performance, energy efficiency, and applications.

The National Institute of Computational Fluid Dynamics (CCM) successfully completed the massive parallel computing of the re-entry flight status of the Tiangong-1 spacecraft using the "Unified Algorithm Numerical Simulation for Aerospace Vehicles" software built on the "Sunway TaihuLight." By utilizing 16,384 processors (an eighth of the entire TaihuLight system), they finished a computation that would normally take 12 months in just 20 days, with results aligning well with wind tunnel experiments. Furthermore, materials science, closely linked to the aerospace field, requires simulations to determine the most effective element combinations.

Currently, the peak computing performance of TaihuLight reaches 12.54 petaflops per second. This means that the computational power of one minute of TaihuLight is equivalent to 7.2 billion people using calculators non-stop for 32 years. Clearly, China's supercomputer hardware capabilities are not "empty talk." What then is the true power of such a superpower?

"Intermediate Interruptions" Limit Application Expansion

In 2016, an application project running on TaihuLight won the "Gordon Bell" prize, breaking a 29-year drought for China in high-performance computing applications.

The project, "Ten Million-Core Scalable Global Atmospheric Dynamics Implicit Simulation," utilized over 10 million cores to conduct a numerical simulation of the climate. Its unique feature lies in the algorithm—a novel "implicit" calculation method. This algorithm improves computational efficiency by almost an order of magnitude compared to the current state-of-the-art internationally. Yang Chao, head of the award-winning team and a researcher at the Institute of Software, Chinese Academy of Sciences, stated that the algorithm and performance were the main reasons for the team's success.

Implementing high-performance computing to achieve real-world applications isn't something a single supercomputer can accomplish alone. It's more like a chain, with software connecting the upper application scenarios and lower machine hardware. Professor Yang Guangwen, a professor at Tsinghua University and director of the National Supercomputing Wuxi Center, told a reporter from Science and Technology Daily that researchers cannot design better algorithms and application software if they are unfamiliar with the applications. Similarly, software users lack the necessary computing resources due to their lack of familiarity with the applications. The ability to produce high-quality software... Even when technology is mature and demand is strong, the absence of proper software impedes the practical application of supercomputing.

Moreover, the larger issue China faces is not just the backwardness of software but also a gap in the software ecosystem compared to foreign countries. For instance, Gan Lin noted: The prosperous processor of TaihuLight is akin to coordinating ten million ants pulling heavy objects simultaneously. If one ant "falls," the load needs to be redistributed to other software. This seemingly simple step requires common software to manage and form mechanisms for fault tolerance and error correction. While China has these capabilities, they are not as mature as those in foreign software. Most programs running on TaihuLight required code modifications to optimize the hardware characteristics of TaihuLight. The tuning process depends heavily on software tools.

"Even though we've won the 'Gordon Bell' award, there is still a long way to go to fully enhance our country's supercomputing application level," Professor Yang Guangwen stated bluntly. This is similar to taking a newly developed product out of the lab—its effectiveness is predictable, but practical implementation remains to be tested. Professor Yang believes that researchers need to possess an "artisan spirit" to excel in supercomputing.

From hardware to software, China has been catching up and making continuous efforts to close the gap. At the national level, several projects within the major national scientific and technological special projects are being funded under the support of supercomputing, with investments steadily increasing. Professor Yang Guangwen noted that the system's policies, projects, and talent support will spur the development of ultrahigh-performance R&D and nurturing industries, promoting national industrial transformation and upgrading.

There is no competitive relationship between quantum computing and hypercomputing.

At the "2017 International Big Data and E-Class Computing Symposium," Professor Yang Guangwen revealed that the development of the next-generation exascale supercomputers based on "Sunway TaihuLight" has been included in the national "13th Five-Year Plan" and is currently progressing rapidly. The focus is on developing three prototypes—"Sunway," "Tianhe," and "Dawning." It is expected that the first domestically produced exascale supercomputer will be launched around 2020.

Scaling from billions to quintillions presents immense challenges. Professor Yang told the Science and Technology Daily reporter that the increase in computing power has led to rising power consumption. Addressing this involves designing low-power chips, adopting efficient cooling technologies, and optimizing the entire system's management—all issues currently being tackled.

Compared to emerging quantum computers, Ms. Lin believes that the next generation of supercomputers will not compete with them for decades. Traditional computers represent one bit as either 0 or 1, whereas quantum computers rely on quantum uncertainty to enable polymorphic representations. For example, a qubit can exist in multiple states, potentially making quantum computers far more powerful than traditional ones at the same scale. However, the uncertainty in molding time adds more uncertainty to this potential.

"For quantum computers to be mature, they require a mature algorithm that can simulate applications and match the quantum uncertainty to produce the desired results," said Gan Lin, which also limits their applicability to specific domains.

"In the next generation of supercomputing development and competition, China must not only win the battle of speed but, more importantly, win the battle of application!" Professor Yang emphasized. The lagging domestic application software and shortage of talent have severely hindered the development of supercomputing. With the ultimate goal of "application," the "Sunway TaihuLight" team will further integrate domestic and international strengths to jointly create a domestic chip ecosystem and strengthen application software development, particularly in major application areas such as weather and climate services, advanced manufacturing, life and health, and big data analysis. They aim to address application bottlenecks and make up for the shortcomings in Chinese supercomputing applications.

FUME

FUME EXTRA 1500PUFFS

Shenzhen Xcool Vapor Technology Co.,Ltd , http://www.xcoolvapor.com