At the Huawei Connect 2025 event in Shanghai last week, Chinese technology Huawei explained its plans for the next generation of its Ascend Chip line and published details of its deployment in and world -class leading clusters.
At the Huawei Connect Connect conference, Vice-Chairman of the Board of Directors of the Board of Directors of Huawei, Eric Xu, he said that 2025 was an unforgettable year, and claimed that Deepseek-R1’s debut in January was a turning point for Huawei. He acknowledged what industry commentators had previously confirmed; China is likely to lag in the processing of the semiconductor, “for a relatively long time”.
Due to tariffs, politically motivated rhetoric and commercial embargoch, Huawei will remain progress in the field of infrastructure and related technologies and produce powerful clusters associated with an open -end protocol. Several large pieces of previously proprietary software, including the AI OpenPang Foundation and its Mind SDK, decided.
New range of ascend chip
The company plans to produce three new series Ascend Chip, which calls 950, 960 and 970.
The Ascend 950PR and 950o will provide additional support for low accurate data formats, including FP8: The 950 provides power and MXFP8 delivers two PFLOPS. (Pflop is a thousand trillion calculations with a moving point per second.)
The company says the 950 will offer better processing of the vector and will be able to granular access to memory, processing of 128 bytes, out of 512 bytes.
The Ascend 950 chips will offer 2 TB/s interconnect bandwidth, two and a half more than the current Ascend 910C, which in May this year entered production for Huawei customers. 950PR will be available in the first quarter of 2026, with Ascend 950DT after Q4.
A year later in the fourth quarter of 2027, the Ascend 960 will offer twice the computing force, the width of the memory access, memory capacity and the number of connecting ports such as its Stablete, 950.
The end of 2028 will see the establishment of the most respected Huawei chip, Ascend 970. Xu said the conference: “We are still working on some of its specifications, but our general goal is to move all our specifications much higher.” He said that the Ascend 970 series will be able to communicate via the 4TB/s bandwidth, burn 8 PFLOPS FP4 and will have an even more memory capacity than the previous Ascend range.
(Eric Xu, Huawei
Superpods NPU
The Huawei strategy is rather than focusing on the production of chips that are stronger than its main competitor, Nvidia, to offer hyperscalers clumps of finished calculation in the form of superpods. These will begin to appear Q4 2026 in the form of SuperPod Atlas 950 loaded by Ascend 950DT chips.
The competing NVIDIA NVL144 system (closest competitor SuperPod) is to be launched in the middle to the end-late-2026. Huawei claims that its first superpod is almost seven times higher processing performance and has 56.8 times more NPUs than the NVL144 GPU number. NVIDIA NVL576, set up to 2027, will still be a portable superpatient atlas 950, says Huawei.
General Computer Science (unmistakable) series of chips
For general computer techniques, Huawei plans to release two models of its Kunpeng 950 processors at the beginning of 2026. They will include 96 cores and 192 fibers or 192 cores and 384 fibers in a faster model. At the same time, Huawei plans to publish what the XU called the “first general superputor in the world”, the Taishan 950 based on the Kunpeng 950.
Open Source Code connection protocol, Unifiedbus 2.0
The NPU and “General Computing” will use Huawei Unifiedbus 2.0, which is an update of the current Unifiedbus 1.0. Version 1.0 Unifiedbus is a connection technology used by the Atlas 900 A3 Superpod, which has appeared in March this year, and has more than 300 institutions in DC and institutions around the world.
Unifiedbus 2.0 will have a license as an open-source and the technical specifications are to be released immediately into the developer community. Unifiedbus 2.0 cuts their teeth in new generations of superpodes and will be a superstructure cluster protocol, inevitably named Superklusters.
The first SuperCluster will be the SuperCluster Atlas 950, which will offer 2.5 times more NPU and 1.3 times more computing power than the Xai’s Colossus cluster, currently the strongest in the world.
At the end of 2027, Huawei SuperCluster Atlas 960, which will include more than a million NPU and deliver 4 ZFlops in FP4^21 Operations with a moving part per second) through a single virtualized computing instance. “Superlodes and super -packed unifiedbus are our answer to the sharp demand for calculation, both today and tomorrow,” Xu said.
(The source of the picture: “Ejection of coronal materials in the sun” from deceitful views is licensed under CC 2.0.)
Do you want to learn more about data centers from the leaders of the industry? Take a look at the Data Center Congress, which takes place in Amsterdam, California and London. The complex event is part of TechEx and together with other leading technological events. Click here for more information.
Techforge Media is powered by News. Explore other upcoming business technology and webinars here.
(Tagstotranslate) ascend