Huawei discontinues open-source AI toolkit availability for developers following China's pressure on NVIDIA
Huawei, a leading Chinese tech company, has announced the formation of the Model-Chips Ecosystem Innovation Alliance, a move aimed at building an open-source ecosystem around its Compute Architecture for Neural Networks (CANN) platform. This strategic move comes amid increased scrutiny of NVIDIA's role in China's AI infrastructure.
The CANN platform, first introduced in 2018, is a software stack designed to support Huawei's AI chips and serve as an alternative to NVIDIA's dominant CUDA platform in AI development. It enables running neural network computations efficiently on Huawei’s Ascend AI processors, forming part of Huawei’s self-built AI stack that includes chips, software, data centers, and specialized AI models like the Pangu series.
Huawei's challenge to NVIDIA’s dominance primarily revolves around building a parallel AI ecosystem, focusing on industry-specific AI models, achieving hardware performance parity, enabling CUDA compatibility, and capitalizing on geopolitical factors.
Unlike NVIDIA’s CUDA, which is the leading global AI software ecosystem, CANN is tailored for Huawei’s hardware and the Chinese AI market, with growing presence globally, especially in Belt and Road countries. Huawei’s flagship Pangu AI models are not broad chatbots like ChatGPT but targeted at specific sectors such as mining, healthcare, finance, and government, where these models have practical deployments.
Huawei’s Ascend 910B/C/D chips rival NVIDIA’s latest AI GPUs, and their CloudMatrix 384 supernode cluster outperforms NVIDIA’s flagship GB200 NVL72 in raw throughput and memory bandwidth, albeit with lower power efficiency offset by lower electricity costs in China. To overcome developer resistance against adopting the less mature and somewhat unstable CANN framework, Huawei has introduced intermediary software that translates CUDA commands to its own chips, allowing existing AI software written for CUDA to run on Huawei hardware without extensive rewriting.
The formation of the alliance comes amid increased scrutiny of NVIDIA's role in China's AI infrastructure. The Cyberspace Administration of China (CAC) has launched an inquiry into NVIDIA's H20 GPU, a GPU designed by NVIDIA specifically for China to comply with Washington's export restrictions. US experts have claimed that remote-control technologies related to NVIDIA chips have matured, and there are growing concerns among US lawmakers, who have urged NVIDIA to embed tracking features into advanced chips.
The alliance, which includes Huawei's Ascend unit, Tencent-backed StepFun, Infinigence AI, SiliconFlow, MetaX, Biren Technology, Enflame, Iluvatar Corex, Cambricon Technologies, and Moore Threads, aims to push the adoption of domestic AI chips and create an alternative to NVIDIA's proprietary CUDA software stack. This move supports China's goal of tech self-reliance, spanning computing hardware, software, and services, and aligns with Beijing's broader push for technological self-sufficiency.
References:
- Huawei Unveils AI Strategy to Challenge NVIDIA
- Huawei Challenges NVIDIA's AI Dominance with CANN
- Huawei's CANN Platform: A Challenge to NVIDIA's Global Leadership
- Huawei's CloudMatrix 384 Outperforms NVIDIA's Flagship AI Supercomputer
- The Model-Chips Ecosystem Innovation Alliance, formed by Huawei, seeks to build an open-source ecosystem around its CANN platform, aiming to challenge NVIDIA's dominance in AI development and cyberscurity, especially in the areas of science, technology, and finance.
- With the growing presence of CANN, tailored for Huawei’s hardware and the Chinese AI market, there's a promising potential for innovation, as the alliance strives to create a parallel AI ecosystem and promote the use of domestic AI chips as an alternative to NVIDIA's proprietary CUDA software stack.
- The alliance's focus on industry-specific AI models, like those developed by Huawei in sectors such as mining, healthcare, finance, and government, could lead to groundbreaking advancements in science and technology, contributing significantly to China's broader push for technological self-sufficiency and tech self-reliance.
