The Shift in AI Ecosystems: Huawei vs. NVIDIA
When discussing the landscape of AI hardware, the spotlight often shines on NVIDIA’s achievements: the H100, Blackwell, energy consumption, and manufacturing processes. However, this focus overshadows a critical aspect of NVIDIA’s competitive edge: CUDA.
The Significance of CUDA
CUDA is not merely a feature; it is the de facto standard for AI development. Most AI code is crafted, optimized, and debugged within CUDA’s framework. Switching GPUs without changing CUDA is virtually impossible, as it necessitates the rewriting of extensive codebases developed over years. This barrier makes CUDA a formidable defensive moat for NVIDIA, reinforcing its position in the AI ecosystem.
Huawei’s Ambitious Strategy
Huawei’s primary goal is not to simply create a counterpart to NVIDIA’s hardware, like a Chinese H100. Instead, its focus is on establishing a pathway that allows developers to transition to its Ascend chips without a dramatic paradigm shift. Importantly, the ongoing export restrictions imposed on Huawei are accelerating its efforts to cultivate this alternative ecosystem.
A Divided Ecosystem
The current state of the AI landscape can be summarized in two camps:
- An ecosystem centered around NVIDIA.
- A competing ecosystem being developed by China, led by Huawei.
In this context, Huawei’s approach goes beyond chip manufacturing; it aims to build an expansive ecosystem across AI and other sectors. The possibility remains that Huawei could lag in hardware but still foster dependency through software.
Three-Pronged Approach to Ecosystem Development
Huawei’s strategy can be broken down into three core fronts:
Native Stack (CANN + MindSpore): This is Huawei’s pure alternative, focused on creating an environment optimized for Ascend. However, challenges such as high costs, instability, incomplete documentation, and a smaller community hinder its widespread adoption.
PyTorch Support: This strategic move seeks to retain the existing developer community by minimizing the need to abandon PyTorch. The
torch_npuadapter allows developers to run their PyTorch models on Ascend. However, this solution is not natively integrated, leading to potential performance issues with each PyTorch update.Portability via ONNX: Huawei recognizes the potential for inference and deployment rather than just training. By utilizing ONNX as a bridge format, developers can train models on compatible platforms and deploy them on Ascend, offering immediate solutions if shortages occur.
The Tactical Advantage
Huawei’s methodology mirrors the tactics that propelled NVIDIA to its current stature: transforming hardware into a holistic user experience. To facilitate the transition, Huawei places engineers directly within client environments to assist with code migration and optimization. While this might not scale as a long-term business model, it serves to buy crucial time for Huawei to mature its tools and support systems.
Building Infrastructure Over Time
Should Huawei manage to integrate enough teams into the Ascend ecosystem, these changes could evolve from necessity to habit and ultimately become infrastructure. This shift wouldn’t necessarily indicate superiority; rather, it could stem from basic integration into the industry.
Challenges Ahead for Huawei
Despite its advancements, Huawei faces two significant hurdles:
Hardware Improvement Rate: Current analyses suggest that Huawei’s hardware development is relatively stagnant. If NVIDIA continues to expedite its cycles, the gap could widen further.
Off-Chip Bottlenecks: Challenges related to memory (HBM), tools, and industrial capacity remain pertinent. While Huawei may produce “worse” chips, the sheer volume needed poses logistical challenges.
The Road Ahead
If these trends persist, we can expect to see two notable outcomes:
- A decline in chip hype and increased reporting on real migration stories, detailing how teams transition to Ascend and the challenges they encounter.
- A shift away from solely training in Ascend toward normalizing a hybrid approach, where developers train on the most accessible hardware and deploy where necessary.
In conclusion, NVIDIA’s future will continue to hinge on CUDA, representing more than just a chip—it is a strategic escape route. With the current geopolitical landscape, Huawei may not just challenge NVIDIA but redefine the AI ecosystem in the process.