Huawei is reportedly making a major strategic shift in its approach to GPU design, aiming to build next-generation hardware that serves a broad range of computing needs—not just artificial intelligence. This move could position the Chinese tech giant as a direct competitor to industry leaders like NVIDIA, especially as U.S. export restrictions continue to reshape the global AI chip market.
Ascend 920: From AI Accelerator to General-Purpose GPU
Huawei’s upcoming Ascend 920 chip is at the center of this transformation. Unlike previous generations, which were tailored primarily for AI workloads, the Ascend 920 is being designed as a multi-purpose GPU. This means it will be capable of handling a wide spectrum of data processing and computing tasks, moving beyond the confines of specialized AI acceleration.
The Ascend 920, built on SMIC’s 6nm process, is expected to deliver over 900 TeraFLOPS of BF16 half-precision performance and a memory bandwidth of 4,000 GB/s using HBM3 modules. These specs represent a 30-40% performance boost over the current Ascend 910C, which already powers Huawei’s CloudMatrix AI clusters. The new chip will also support PCIe 5.0 and advanced interconnect protocols, improving integration in large-scale server deployments.
Competing with NVIDIA in China
Huawei’s pivot comes as U.S. export bans have cut off Chinese access to NVIDIA’s most advanced AI chips, such as the H20 and H100. These restrictions have left a significant gap in the Chinese market, which Huawei aims to fill with its domestically produced Ascend 920. The chip is positioned as a direct alternative to NVIDIA’s H20, with real-world benchmarks suggesting it could outperform the U.S.-made rival in some scenarios, albeit with higher power consumption.
Software Ecosystem: Toward CUDA Compatibility
A key challenge for Huawei has been the software ecosystem. While NVIDIA’s CUDA platform dominates GPU programming for AI, Huawei’s chips currently rely on its proprietary CANN (Neural Network Computing Architecture). Reports suggest that Huawei is developing middleware to enable CUDA compatibility, allowing developers to run existing CUDA-based applications on Ascend hardware with minimal changes. This could dramatically lower the barrier for adoption and make Huawei’s GPUs more attractive to a global developer base.
Broader Ambitions and Industry Impact
Huawei is also reportedly exploring hardware features modeled after both NVIDIA and AMD GPUs, signaling a commitment to aligning with global industry standards. While the Ascend 920 is not expected to compete with NVIDIA’s top gaming GPUs, its enhanced versatility could make it a compelling choice for AI developers and enterprises seeking alternatives to U.S.-controlled hardware.
Mass production of the Ascend 920 is expected to begin in the second half of 2025, with the chip likely to play a central role in China’s efforts to achieve technological self-sufficiency in advanced computing.
“Huawei’s new Ascend 920 AI chip, built on SMIC’s 6nm process, delivers 900 TFLOPs and 4TB/sec memory bandwidth, outperforming its predecessor and rival NVIDIA’s restricted H20 GPU. Integrated in the CloudMatrix AI server, it offers 30-40% higher performance, positioning Huawei as a strong competitor in AI hardware.”
As the AI hardware landscape evolves, Huawei’s push for a general-purpose GPU could reshape the competitive dynamics—and potentially offer a new path forward for developers and enterprises navigating a rapidly changing market.