Recently, Google’s TPU has gained significant attention, and Huawei has adjusted its chip technology roadmap. This shift is driven by the evolving demand for AI computing power and the changing scenarios for technology adaptation. First, let’s clarify the similarities and differences between NPU and TPU, and then examine the logic behind Huawei’s choice:
1. NPU vs TPU: Technical Essence and Design Goals
(1) Similarities: Designed for AI Computing Acceleration
Both belong to “Domain-Specific Architectures (DSA)” focused on AI tasks (such as training/inference of deep learning), enhancing AI computing efficiency and reducing power consumption through customized hardware architecture (as opposed to the “one-size-fits-all” logic of general-purpose CPUs/GPUs). The core is the “hardware-level adaptation” to AI algorithms (such as neural networks): for instance, designing dedicated computation units for high-frequency operations like matrix multiplication and activation functions, thereby reducing the overhead of general-purpose chips.
(2) Differences: Architectural Direction and Ecological Focus
| Dimension | NPU (e.g., Huawei’s early Ascend) | TPU (Google’s Cloud-Native Chip) |
|---|---|---|
| Design Origin | Targeting edge + some cloud AI inference (e.g., real-time recognition in smartphones and cameras), balancing low power consumption and lightweight computing. | Focused on large-scale cloud training (Google Search, DeepMind model training), pursuing extreme computing density and cluster efficiency. |
| Architectural Focus | Optimized for “inference scenarios”: tighter circuits, memory scheduling adapted for small data volumes and low latency; early versions had weak support for training (precision and bandwidth bottlenecks). | Full-stack support for “training + inference”: designed for ultra-large models (e.g., Gemini), emphasizing high-bandwidth memory and large computing cluster interconnects, supporting training at the trillion-parameter level. |
| Ecological Binding | Closed-loop within Huawei’s ecosystem (HarmonyOS devices, Ascend servers), adapting self-developed frameworks (MindSpore). | Deeply integrated with Google Cloud ecosystem (GCP) and TensorFlow framework, forming the core of Google’s AI infrastructure as a “cloud + chip + framework” trinity. |
2. The Logic Behind the Popularity of Google’s TPU
The recent attention on TPU is fundamentally driven by the “arms race” in the AI industry for large models, necessitating an upgrade in computing infrastructure:
- Essential Need for Large Model Training: Training models with hundreds of billions/trillions of parameters requires extreme computing density and cluster efficiency—TPU was born specifically for “large-scale cloud training”. Google uses TPU clusters to support Gemini and DeepMind development, validating its overwhelming efficiency in training scenarios.
- Cloud Vendor Arms Race: AWS (Trainium), Azure (ND-H100), and Google (TPU v5e/v6) have all launched dedicated AI chips, with TPU serving as a visible symbol of “Google’s AI technology moat”, naturally becoming a focal point in the industry.
- Open Source + Ecological Spillover: Google binds TPU to the TensorFlow ecosystem, attracting numerous research institutions and startups to develop based on TPU, further amplifying its popularity.
3. The Underlying Logic Behind Huawei’s Shift to GPGPU Solutions
Huawei is not “abandoning NPU”; rather, it is expanding its technology roadmap in response to the diversification of AI computing needs. The core drivers are:
- Changing Scene Requirements:
– In the early years, NPU primarily targeted edge inference (e.g., AI optimization for smartphone photography, edge device recognition), but the industry’s focus has now shifted to “large model training + general AI computing”—training requires “parallel computing power that can support trillion-parameter levels across cards and clusters”, while traditional NPU architectures (especially those optimized for edge) are inherently limited in memory bandwidth, multi-card interconnects, and precision support.
– Enterprise customers (e.g., finance, industry) require a mix of “general + specialized” computing power: they need NPU for inference acceleration and GPGPU to handle traditional scientific computing (e.g., fluid dynamics simulation) + emerging AI training. The maturity of the GPGPU ecosystem (CUDA ecosystem) is higher, with broader scene adaptability. - Ecological Competition Reality:
– The NPU ecosystem (e.g., Huawei MindSpore + Ascend NPU) is relatively closed, while GPGPU relies on NVIDIA’s CUDA to form a mature ecosystem of “framework (PyTorch/TensorFlow) + toolchain + developer community”, resulting in lower migration costs for enterprises. Huawei’s development of GPGPU (e.g., Ascend AI chips compatible with GPGPU mode or developing similar architectures) is a pragmatic choice to “integrate into the mainstream AI ecosystem”. - Completing the Technology Roadmap:
Huawei’s “GPGPU layout” is not a replacement for NPU, but rather a combination of “NPU guarding the edge inference base + GPGPU attacking cloud training + general computing power”—after all, in the era of large models, “training computing power” is the strategic high ground of the AI industry, and GPGPU is currently the mainstream hardware form supporting training (with NVIDIA H100 and others accounting for over 80%). Huawei must complete this capability.
4. Conclusion: Technology Roadmaps Always Serve Demand
NPU and TPU are products of the “division of labor in AI computing power”: NPU is suitable for lightweight inference on the edge, while TPU specializes in large-scale cloud training. Huawei’s expansion into GPGPU is due to the industry’s demand shifting from “edge inference” to “training + general computing power + ecological compatibility”; the technology roadmap must evolve with demand. Google’s TPU has exploded in popularity due to the “essential need for large model training” colliding with “Google’s full-stack advantages in AI infrastructure”; Huawei’s adjustment is a strategic adaptation to the “changing phases of the AI industry”—both are essentially inevitable choices matching “computing power supply” with “industry demand”.