.jpg "It Still Feels Like Summer in Parts of the US. Blame El Niño")
The Chinese tech giant Huawei has said it is seeing strong domestic demand for its chips to reduce reliance on Western hardware and unveiled a next-generation chip intended to compete with the likes of Nvidia.
At the annual Huawei Connect conference, held in Shanghai, chairman Eric Xu said that Huawei was pushing for its chips to meet the performance demands of China’s domestic AI labs and that its end goal was total self-dependency.
“We cannot accept a destiny where we cannot control our fate being determined by others in terms of willingness to sell chips to China or not,” he said.
“No matter if it’s for the Chinese government, industry in China, or for Huawei, it is certainly the way forward to try to push for full self-sufficiency for chips.”
Huawei’s current flagship chip is the Ascend 950DT, which is specialised for AI and capable of delivering nearly 2TB/s of memory interconnect bandwidth and a petaflop of 8-bit floating point (FP8) peak compute power. FP8 is the low-precision data type often used for AI inference.
The Ascend 960DT, an AI training focused chip which Huawei said will be available in the first quarter of 2027, is intended to double this power. This chip will be followed by the Ascend 960PR, optimised for massive model inference and expected in the third quarter of 2027.
In addition to raw power improvements the Ascend 960 chips will introduce HiF4, Huawei’s proprietary approach to 4-bit floating point (FP4) operations. This is key as inferencing AI models at FP4 allows operators to double the tokens per second output by AI models, as well as reduce operating costs by running models in a lower memory, compressed state.
At its event, Huawei also announced the Atlas 960 SuperPoD, a computing cluster formed from up to 4,096 Ascend 960 chips that Huawei said will deliver up to eight exaflops of FP8 compute performance.
This is the equivalent of just over six Nvidia Rubin NVL72 server racks, or around 18 per cent the peak performance of a Google Cloud TPU v7 Ironwood cluster.
Huawei acknowledged that frontier AI models may have 10 trillion parameters or more and that individual clusters would not be enough to train models on this scale. In response it has developed SuperCluster, hundreds of SuperPoDs connected using its new Peerium Computing Architecture to allow for up to a million Ascend chips to work in tandem on a training run.
Huawei has benefitted from US bans on exporting advanced AI chips to China, as well as moves by Beijing to protect its domestic technology and defend the profits of its sovereign chip designers.
In May, Huawei said its chip sales will grow by at least 60 per cent in 2026, amid strong demand for the Ascend 950PR.
Going forward, Xu indicated, Huawei will release a new chip family each year. On this timetable, the Ascend 970 and 980 models will release in 2028 and 2029.
