Huawei Ascend C Programming Language launched, compatible with C/C++

During the 2023 Kunpeng Ascend Developer Summit, Huawei launched new software and hardware technologies and solutions. Among these new announcements, Huawei unveiled a brand new Ascend C Programming language.

According to Huawei, the Ascend C programming language is mainly designed for operator development scenarios. And it natively supports C/C++ programming specifications and matches the user development environment.

Structured kernel feature programming and the core logic of the operator can be adapted using the high-performance class library interface. Meanwhile, the automatic pipeline parallel scheduling realizes the best execution performance of the operator.

It also has a Kernel-level source code migration tool for GPU, which can reduce operator migration time from “week level” to “day level”.

Related – Huawei Kunpeng and Ascend AI exceeds 3.5 million developers

For large models, Ascend leverages its comprehensive advantages in software and hardware to accelerate large model innovation throughout the entire process. This includes training development, and inference deployment and can support 25-day stable training of large models with 200 billion parameters.

Zhang Dixuan, President of Huawei Ascend Computing Business, said that currently over 30 mainstream large-scale models have been incubated and adapted based on Ascend. More than half of China’s original large-scale models have been incubated based on Ascend AI, including the recently launched Xunfei Spark.

Zhou Jun, President of Huawei ICT Marketing, said that Shengteng has achieved 1100+ partners. As of now, 25 cities have built artificial intelligence computing centers based on Shengteng and 14 of them are now fully operational.

The Chinese tech maker promised that Huawei will focus on root technology, and continue to strengthen the Ascend AI industry ecosystem. It will implement the open source strategy and contributes to the Ascend open source community, and make the Ascend MindSpore the preferred framework for supporting AI innovations such as large models and scientific intelligence.