Chinese language know-how large Huawei introduced its plans for the subsequent generations of its Ascend chip line on the Huawei Join 2025 occasion in Shanghai this week.
In his keynote to the convention, deputy chair of the Huawei board, Eric Xu, stated that 2025 had been a “memorable 12 months,” and famous the debut of DeepSeek-R1 in January as a turning level for the corporate. He additionally acknowledged that China is prone to lag behind in semiconductor manufacturing course of nodes, “for a comparatively very long time.”
The corporate’s response to tariffs and commerce embargoes is to advance infrastructure design and know-how, plus it’s made the choice to open-source a number of giant swathes of its software program, together with the openPangu basis AI fashions and the Thoughts sequence SDKs.
The brand new Ascends
The corporate plans to produces three new sequence of the Ascend chip, the 950, 960, and 970.
The Ascend 950PR and 950TO shall be forged from the identical die, and supply extra assist for low-precision information codecs, together with FP8 – the place the 950 will ship a PFLOP of efficiency, and MXFP8, rated at two PFLOPs. A PFLOP is one thousand trillion flotaing level calculations per second.
There’ll even be higher vector processing, and extra granular reminiscence entry, all the way down to 128 byte chunks from 512 bytes.
The Ascend 950 chips will provide 2 TB/s interconnect bandwidth, 2.5x greater than the present Ascend 910C. The 950PR shall be accessible Q1 2026, and the Ascend 950DT launches This autumn 2026.
Obtainable a 12 months later in This autumn 2027, the Ascend 960 may have twice the computing energy, reminiscence entry bandwidth, reminiscence capability, and variety of interconnect ports because the 950. It is going to assist Huawei’s proprietary HiF4 information format, which, the corporate claims, brings better precision than different FP4 applied sciences.
Probably the most succesful chip would be the Ascend 970, slated for launch This autumn 2028. Xu stated, “We’re nonetheless engaged on a few of its specs, however our common purpose is to push all of its specs a lot increased.” He stated it was anticipated that the Ascend 970 sequence will provide an interconnect bandwidth of 4TB/s, be able to 8 PFLOPs of FP4, and can include bigger reminiscence capability.
SuperPods of NPUs
Huawei’s technique is to supply hyperscalers clusters of uncooked compute within the type of SuperPoDs, which can seem start to seem This autumn 2026 within the type of the Atlas 950 SuperPoD, geared up with the brand new Ascend 950DT chips.
Competitor NVIDIA’s NVL144 system (a SuperPod analogue) will launch mid- to late-2026, and Huawei claims that its first SuperPoD may have 56.8 instances extra NPUs than GPUs within the NVL144, and ship practically seven instances the processing energy. Even with the scheduled arrival of the NVL576, which NVIDIA is ready to launch in 2027, the Atlas 950 SuperPoD will nonetheless be the higher performer.
Normal computing chips
For common computing, Huawei plans to launch two fashions of its Kunpeng 950 processors in Q1 2026, sporting 96 cores & 192 threads, and 192 cores & 384 threads within the sooner of the 2 fashions. There can even be what Xu referred to as “the wold’s first general-purpose computing SuperPoD,” the Kunpeng 950-based TaiShan 950 SuperPod, which shall be accessible within the first quarter of 2026.
Open-source connectivity protocol
The NPU and common computing SuperPoDs will use UnifiedBus 2.0, the subsequent iteration of the prevailing UnifiedBus 1.0. That’s the interconnection know-how utilized by the Atlas 900 A3 SuperPoD, which got here into service in March this 12 months, with over 300 installations thus far.
UnifiedBus 2.0 is to be an open protocol, with the tech specs launched instantly to developer neighborhood. UnifiedBus 2.0 shall be used internally within the new generations of SuperPods, and join clusters of SuperPods, forming SuperClusters.
The primary cluster product is to be the Atlas 950 SuperCluster, providing 2.5 instances extra NPUs and 1.3 instances extra computing energy than xAI’s Colossus, presently the world’s strongest computing cluster.
Within the final quarter of 2027, Huawei intends to launch the Atlas 960 SuperCluster, which can combine over 1,000,000 NPUs and ship 4 ZFLOPS in FP4 (with a ZFLOP representing 10^21 floating level operations per second). “SuperPoDs and SuperClusters powered by UnifiedBus are our reply to surging demand for computing, each right this moment and tomorrow,” Xu stated.
(Picture supply: Eric Xu, Huawei)
Wish to be taught extra about AI and massive information from business leaders? Take a look at AI & Big Data Expo going down in Amsterdam, California, and London. The excellent occasion is a part of TechEx and co-located with different main know-how occasions. Click on here for extra data.
AI Information is powered by TechForge Media. Discover different upcoming enterprise know-how occasions and webinars here.
