Peking University team ports DFT ABACUS software to Huawei Kunpeng 920 processor, achieving up to 10.7x speedup

Huawei, in collaboration with the China Telecom branch in Shaanxi Province, has commercially deployed 32T32R FDD Massive MIMO technology on the Shu-Lan and Yinxi high-speed railway lines. This equipment operates simultaneously on 1.8 GHz and 2.1 GHz frequencies.

Huawei states that this is the first commercial use of FDD Massive MIMO in a high-speed railway environment in China. High-speed rail networks present significant challenges for mobile coverage due to the large distance between towers, train speeds up to 350 km/h, and penetration losses caused by metal carriages.

Massive MIMO technology, which is often used to increase capacity in densely populated urban areas, forms narrower beams compared to traditional wide-beam antennas. Huawei asserts that this beam shape is suitable for a linear route: the beams can be directed along the tracks, follow the train through cells, and adapt to compensate for the Doppler effect, which could degrade the connection.

The company highlighted three key design aspects for railway application. Adaptive beams are designed to limit frequency shift effects at high speeds. Inter-cell beam tracking and coordinated planning aim to reduce the number of handovers when a train crosses multiple cells. Furthermore, broadcast beams and pilots are configured to ensure linear coverage so that energy remains within the corridor rather than scattering in open terrain.

According to the statement, measurements conducted on the Yinxi section in Shaanxi showed an approximate 200-meter expansion of cell coverage after modernization, a 24% increase in average 5G uplink speed, and an increase in average downlink speed of over 50%. User experience improvements were even more pronounced: uplink speed at the zone boundary increased by 95%, and downlink speed by over 100%.

Although FDD Massive MIMO has long served as a tool for increasing capacity in urban macrocells, deploying 32 transmit and 32 receive chains on a high-speed corridor represents an attempt to apply this urban capacity model to a fast-moving linear cell. Huawei and Shaanxi Telecom present the deployment in Shaanxi as an operational commercial network, not a laboratory experiment. It should be noted that all claimed gains in coverage and throughput are assertions made by the company and operator from the announcement in October and have not undergone independent verification.

The Peng Cheng Laboratory has commenced the test operation of 10,000 petaflops of domestic artificial intelligence computing power at its large-scale scientific site, Peng Cheng Cloud Brain III, located in Shenzhen. This was announced at the opening ceremony of the World Computing Network Conference 2026 (WCNC 2026), held in the Nanshan district of Shenzhen on October 15 and 16, according to the laboratory's official website.

The laboratory stated that this step transitions Cloud Brain III into the 'construction while operating' phase. Currently, 10,000 petaflops of capacity are online, which are already being used to train large domestic models and function as a specialized AI research network for science, while the remaining capacity is still being installed.

According to the Chinese tech publication IT Home, the fully completed system is planned to reach a capacity of 16,000 petaflops, or 16 EFLOPS, and is described as the largest planned AI computing complex in China. Regarding the architecture, the laboratory's announcement only mentioned that the system utilizes an innovative domestic intelligent supernode architecture designed to eliminate data transmission bottlenecks.

IT Home, citing public reports, noted that Cloud Brain III is built on Huawei's Ascend 384 SuperPoD architecture; however, the laboratory itself did not name the chip supplier in its official release. In late December 2025, the laboratory activated the first phase of computational boards with a total capacity of 4,500 petaflops and associated equipment for testing.

Infrastructure Significance and Development Plans

Lu Yutong, Director of the National Supercomputing Center in Shenzhen, told local media that Cloud Brain III provides the city with a solid foundation for inter-district computational connectivity and broad access to computing services. He also added that high-speed connections will extend to Hong Kong and Macau. This facility is designed as a key node of the China Computing Network (C2NET), a national pilot program that, according to the laboratory director Gao Wen, has been underway for four years.

In his main presentation, Gao outlined the following goal, named ZPG: zettaflop-scale computing, petabit-scale communication speed, and gigawatt-scale green energy. He described a three-stage path for the computing network: integrating computing resources, distributing tasks across regions, and ultimately providing computing on demand, similar to electricity.

Two informational documents were also presented at the conference. One, concerning the development of the global computing network, was prepared by the Peng Cheng Laboratory in collaboration with the WCNC alliance. The second, regarding the coordination of computing and energy systems, was created jointly with the Huairou Laboratory and contains a roadmap until 2035, with the period from 2026 to 2030 focusing on planning and foundational establishment, and from 2031 to 2035 focusing on coordinated operation and system optimization.