On the Huawei Join 2025 occasion in Shanghai final week, Chinese language know-how Huawei defined its plans for the following generations of its Ascend chip line, and launched particulars on its deployment in pods, and world-leading compute clusters.
Talking on the Huawei Join convention, deputy chair of Huawei board, Eric Xu, stated 2025 had been a memorable yr, and claimed the debut of DeepSeek-R1 in January was a turning level for Huawei. He acknowledged what trade commentators have beforehand confirmed; China will likely be more likely to lag behind in semiconductor manufacturing processing, “for a comparatively very long time.”
Within the face of tariffs, politically-motivated rhetoric, and commerce embargoes, Huawei will stay forward by advancing infrastructure design and related applied sciences, producing highly effective clusters related by an open-sourced protocol. It’s additionally determined to open-source a number of giant items of beforehand proprietary software program, together with the openPangu basis AI fashions and its Thoughts SDKs.
The brand new Ascend chip vary
The corporate plans to produces three new sequence of the Ascend chip, which it’s calling the 950, 960, and 970.
The Ascend 950PR and 950TO will present extra assist for low-precision information codecs, together with FP8: the 950 mannequin will ship a PFLOP of efficiency, and at MXFP8, ship two PFLOPs. (A PFLOP is one thousand trillion floating level calculations per second.)
The corporate says the 950 will supply higher vector processing, and be able to granular reminiscence entry, processing 128 byte chunks, down from 512 bytes.
The Ascend 950 chips will supply 2 TB/s interconnect bandwidth, two-and-a-half greater than the present Ascend 910C, which went into manufacturing for Huawei clients in Might this yr. The 950PR will likely be accessible within the first quarter of 2026, with the Ascend 950DT following This fall.
A yr later in This fall 2027, the Ascend 960 will supply twice the computing energy, reminiscence entry bandwidth, reminiscence capability, and variety of interconnect ports as its stablemate, the 950. It would assist Huawei’s proprietary HiF4 information format, which brings better precision than different FP4 applied sciences, the corporate says.
The top of 2028 will see the emergence of Huawei’s most succesful chip, the Ascend 970. Xu instructed the convention, “We’re nonetheless engaged on a few of its specs, however our basic objective is to push all of its specs a lot larger.” He stated the Ascend 970 sequence will be capable of talk through interconnect bandwidth of 4TB/s, burn by way of 8 PFLOPs of FP4, and can include a fair bigger reminiscence capability than earlier iterations of the Ascend vary.
(Eric Xu, Huawei
SuperPods of NPUs
Reasonably than concentrating on producing chips which are extra highly effective than its major competitor, Nvidia, Huawei’s technique is to supply hyperscalers clusters of ready-made compute within the type of SuperPoDs. These will start to look This fall 2026 within the type of the Atlas 950 SuperPoD loaded with Ascend 950DT chips.
Competitor NVIDIA’s NVL144 system (the closest SuperPod competitor) is about to launch mid- to late-2026. Huawei claims that its first SuperPoD will ship practically seven instances its processing energy, and have 56.8 instances extra NPUs than the variety of GPUs within the NVL144. Nvidia’s next-generation NVL576, set to be launched in 2027, will nonetheless be out-performed by the Atlas 950 SuperPoD, Huawei says.
Basic computing (non-AI targeted) chip vary
For basic computing, Huawei plans to launch two fashions of its Kunpeng 950 processors firstly of 2026. They’ll comprise of 96 cores & 192 threads, or 192 cores & 384 threads within the quicker mannequin. On the similar time, Huawei plans to launch what Xu known as, “the wold’s first general-purpose computing SuperPoD,” the Kunpeng 950-based TaiShan 950.
Open-source connectivity protocol, UnifiedBus 2.0
Each the NPU and ‘basic computing’ SuperPoDs will use Huawei’s UnifiedBus 2.0, which is an replace on the present UnifiedBus 1.0. Model 1.0 of UnifiedBus is the interconnection know-how utilized by the Atlas 900 A3 SuperPoD, which appeared in March this yr, and already has greater than 300 installations in DCs and establishments all over the world.
UnifiedBus 2.0 will likely be licensed as open-source, and the technical specs are to be launched instantly to the developer neighborhood. UnifiedBus 2.0 will lower its tooth within the new generations of SuperPods, and be the connection protocol for clusters of SuperPods, inevitably named SuperClusters.
The primary SuperCluster would be the Atlas 950 SuperCluster, which is able to supply 2.5 instances extra NPUs and 1.3 instances extra computing energy than xAI’s Colossus cluster, at present the world’s strongest.
In the direction of the top of 2027, Huawei will launch the Atlas 960 SuperCluster, which is able to include over one million NPUs and ship 4 ZFLOPS in FP4 (with a ZFLOP representing 10^21 floating level operations per second) through a single virtualised compute occasion. “SuperPoDs and SuperClusters powered by UnifiedBus are our reply to surging demand for computing, each right now and tomorrow,” Xu stated.
(Picture supply: “Coronal Mass Ejection on the Solar” by Blatant Views is licensed below CC BY 2.0.)
Need to be taught extra about Information Centres from trade leaders? Try Data Centre Congress going down in Amsterdam, California, and London. The great occasion is a part of TechEx and co-located with different main know-how occasions. Click on here for extra data.
CloudTech Information is powered by TechForge Media. Discover different upcoming enterprise know-how occasions and webinars here.
