Let’s put it this way: if we’re talking global scale, Nvidia is the worldwide leader in AI chips manufacturing, but if we’re talking region-based, Huawei is the #1 AI chips manufacturer in China and the Asian continent. However, individually, these companies have fine distinctions in what they do – which is AI chip manufacturing.
As the AI frenzy looms, Nvidia and Huawei have each made names for themselves in the chipset production niche, producing some of the best chips for data‐center training, edge applications, and inference. Nvidia, a long-time global leader, offers high-end GPUs like the H100 and A100, along with specialized inference/edge chips like the L4 and Jetson modules. Huawei, on the other hand, has the Ascend 910 series and Ascend 310, which work best for training and inference, respectively, and power domestic AI clusters and devices.
Here, we would look into the two companies’ portfolios and compare their products, particularly AI chips, in terms of performance, architecture, and ecosystem support. It might be quite a long read. Leggo.
Nvidia vs Huawei AI Chips: The Ultimate Showdown
These two companies are good at what they do, but for now, Nvidia’s flagship chips perform better; Nvidia’s flagship training GPUs leverage cutting-edge process nodes and novel architectures. For example, the H100 uses the Hopper architecture, 80B transistors, TSMC 4N 5nm, and delivers approximately 1,513 FP16 TFLOPS, which doubles to ~3,026 TFLOPS with sparsity, up to 80 GB HBM2e (2.0 TB/s) at 350W.
By contrast, Huawei’s Ascend 910 (first gen) uses the DaVinci architecture, originally 7nm TSMC, and peaks at approx. 256 TFLOPS FP16 (350W). While the second-generation Ascend 910B (SMIC 7nm) improves to roughly 320 TFLOPS FP16 (64GB HBM2e, 400GB/s, 310W), comparable to Nvidia’s A100 (312 TFLOPS) and, by some measures, the Ascend 910B even slightly exceeds the original A100.
However, expert analyses note that Huawei’s 910B variants top out near 400 TFLOPS FP16, which still trails Nvidia’s latest. Huawei now combines two 910B dies into a dual-chip Ascend 910C (8nm SMIC, 53B transistors) to boost throughput: each 910C card can deliver ≈800 TFLOPS FP16. Even so, that is well below the H100’s ≈1,500 TFLOPS per chip.
| Chip (Year) | Vendor | FP16 Throughput | Memory | Power | Notes |
| NVIDIA H100 (2022) | Nvidia | 1,513 TFLOPS (×2,002 TFLOPS FP8 with sparsity) | 80 GB HBM2e @ 2.0TB/s | 350 W | Hopper 4th-gen tensor cores; NVLink/NVSwitch for multi-GPU scaling |
| NVIDIA A100 (2020) | Nvidia | 312 TFLOPS (×624 TFLOPS FP16 with sparsity) | 40/80GB HBM2 (1.6–2.0TB/s) | 400W | Ampere arch; widely deployed in clouds and DGX systems |
| Huawei Ascend 910(A) (2019) | Huawei | 256 TFLOPS | 32GB HBM2 | 350W | DaVinci cores; mixed FP16/FP32/INT8 support |
| Huawei Ascend 910B (2022) | Huawei | 320 TFLOPS (highest models ~400 TFLOPS) | 64GB HBM2e (400GB/s) | 310W | Fewer active cores but added vector units; domestic SMIC 7nm process. |
| Huawei Ascend 910C (2024) | Huawei | ≈800 TFLOPS | 2×64GB HBM2e | ~2×310W | Dual-die (2×910B) package; ~80% of H100 FP16 |
From above, Nvidia’s GPUs offer better performance and also feature very high memory capacity and bandwidth. The H100 offers 80GB HBM2e, supporting trillion-parameter models; Huawei’s chips use older HBM2/e (910B has 400GB/s), limiting single-chip model size. Both companies mitigate this by interconnecting chips: Nvidia’s NVLink/NVSwitch meshes GPUs within servers, while Huawei developed its HCCS interconnect (like NVLink) for Ascend clusters.
Inference and Edge Accelerators
These two giants follow different approaches for AI inference and edge computing chips: Nvidia provides lower-power GPUs like the L4 Tensor Core, A30, and T4, and dedicated SoCs optimized for inference and real-time tasks, while Huawei’s edge/inference line centers on its Ascend 310 and related boards.
Huawei’s Ascend 310 is a tiny 12nm NPU chip capped at 8W power, delivering only 8 TOPS FP16 (16 TOPS INT8). The Nvidia L4 (Ada Lovelace arch.) is a small 1-slot PCIe card (24GB GDDR6) delivering 242 TFLOPS FP16 and 485 TOPS INT8 at just 72W. These modules enable autonomous vehicles, robots, and IoT devices with rich vision/AI capabilities.
The Huawei Ascend 310 integrates an 8-core ARM CPU and a DaVinci AI core for vision and general inference tasks. To increase throughput, Huawei combines multiple 310s on accelerator cards. Huawei also offers Ascend 610 chips for automotive uses, roughly comparable to Nvidia’s Orin CPU+GPU at vehicle scale.
Skip the long talk, Nvidia’s inference chips are generally far more powerful than Huawei’s: the L4 (80W) outperforms multiple Ascend 310s in FP16 throughput. However, Huawei targets ultra-low-power scenarios with 310. Both ecosystems emphasize software stacks, so developers can deploy models efficiently on either platform.
Ecosystem and Adoption
Nvidia’s dominance is as much about software as silicon, and its chips are ubiquitous and even used by Huawei to an extent. Its CUDA/CuDNN platform underpins most AI research and deployment. Popular frameworks like PyTorch, TensorFlow, and many others run natively on Nvidia GPUs, and Nvidia’s inference stack (TensorRT, Triton) is widely adopted by global companies.
Huawei’s ecosystem is mainly China-focused. Its MindSpore deep-learning framework and CANN (Compute Architecture for Neural Networks) toolkit aim to mirror PyTorch/CUDA for Ascend hardware. Chinese tech firms are rapidly adopting Ascend chips; this includes state-backed companies like iFlytek, SenseTime, Baidu, and ByteDance.
Huawei’s stack is still maturing, and basically not engineered for international use, so its current ecosystem is more limited. On the other hand, Nvidia’s stack is widely recognized and adopted globally; Huawei’s software platform has few non-Chinese adopters.
Conclusion
Nvidia is miles ahead of Huawei, but this doesn’t mean Huawei doesn’t have really impressive tech. It’s more about the companies’ target markets; Nvidia focuses on providing global-scale products, while Huawei’s products are mainly state-backed and designed in ways that are favorable to China and Chinese companies. Regardless, year after year, Huawei keeps closing in on the gap it has with Nvidia.
