Huawei Ascend 910

What it is

Ascend 910 is an AI training processor that Huawei launched in Shenzhen on 23 August 2019, together with its MindSpore AI computing framework; it belongs to Huawei's Ascend-Max chipset series.^[1] A training chip carries the heaviest arithmetic in modern machine learning: a network is run forward and backward over data many times, and how quickly the chip does that largely decides how long a model takes to train.

Background

Huawei had set out the family a year in advance. At HUAWEI CONNECT 2018 in Shanghai on 10 October 2018 the company announced its Ascend series of AI chips and their planned specifications, and rotating chairman Eric Xu used the event to present Huawei's AI strategy.^[1] In a September 2025 keynote Xu looked back and dated the first two parts of the line: the Ascend 310 in 2018 and the Ascend 910 in 2019.^[2]

Stated performance

Huawei's published numbers for the 910 were 256 teraflops for FP16 and 512 tera-operations per second for INT8, with a maximum power consumption of 310 W, below the 350 W originally planned.^[1] Teraflops and tera-operations per second both count how many arithmetic operations a chip completes in a second; the difference lies in precision. FP16 work uses 16-bit floating-point numbers, while INT8 uses 8-bit integers, which costs less per operation. Xu said the processor had "more computing power than any other AI processor in the world".^[1]

Huawei also claimed that in a typical ResNet-50 training job, Ascend 910 running with MindSpore was about twice as fast as other mainstream training cards running TensorFlow. That comparison is a company claim rather than an independent benchmark.^[1]

Software around it

The chip was launched as one layer of a larger stack rather than as a standalone part. At launch Huawei said MindSpore would also support GPUs and CPUs, and would be open-sourced in the first quarter of 2020.^[1] The stack also includes the CANN chip enablement layer, which sits between the hardware and the frameworks above it, and the ModelArts platform.^[1]

What it was used for

The most visible large workload built on the hardware was PanGu-α, a family of Chinese language models of up to 200 billion parameters described in a technical report in April 2021. It was developed with MindSpore and trained on a cluster of 2,048 Ascend 910 processors.^[3]

Where it sits in the line

The 910 was not the end of the line. In the September 2025 keynote, Xu recalled that the successor Ascend 910C had become widely known in 2025 through the Atlas 900 A3 SuperPoD, and announced the Ascend 950, 960 and 970 series for the following three years.^[2] Together with the 310 and MindSpore, the 910 is the training layer of Huawei's own AI stack: chips built in-house, a framework to program them, and the training jobs that show what the combination can carry.