Huawei announced the AI accelerator Atlas 350, which outperforms Nvidia H20 in efficiency.

Huawei announced the AI accelerator Atlas 350, which outperforms Nvidia H20 in efficiency.

9 hardware

Huawei launches the Atlas 350 accelerator for AI systems

*Huawei announced the launch of a new accelerator board, the Atlas 350, which is set to become a key component in artificial intelligence solutions. According to the company, it outperforms the NVIDIA H20 adapted for the Chinese market in computational power.*

Technical details
Parameter Value
Chip Ascend 950PR
Power 1.56 petaflops (FP4) – low‑precision format that accelerates data flow
Comparison with NVIDIA H20 2.8× higher

Applications Search recommendations, multimodal generation, and large language models
*The Atlas 350 is a specialized module that can be installed in a server. It is designed to accelerate the inference stages of AI models, including token preprocessing.*

Huawei’s strategy
- Expanding its own chip lineup – new Ascend models are introduced sequentially without using American technology.

- Launch in September 2025 – the Ascend 950PR, focused on preprocessing and recommendations, became part of a three‑year Ascend roadmap.

- Agent AI – growing demand for compute power drives further hardware development.

Storage plans
Huawei intends to massively upgrade its storage lineup in 2026:

- Fully solid‑state solutions OceanStor Dorado and Pacific 9926 for the enterprise segment.

- Deployment of FusionCube A1000 – a compact AI accelerator designed for small and medium businesses.

Storage line president Yuan Yuan emphasized:

> “The first half of the AI era was about compute power. In the second, data is measured. By 2026 Huawei will continue to modernize storage and actively participate in key national projects to build data infrastructure.”

Conclusion:

Huawei is strengthening its presence in AI by offering a more powerful Atlas 350 accelerator based on Ascend 950PR. The company simultaneously expands its own chip lineup and plans major upgrades to storage systems to support the growing demand for compute resources and data for modern agent‑AI solutions.

Comments (0)

Share your thoughts — please be polite and stay on topic.

No comments yet. Leave a comment — share your opinion!

To leave a comment, please log in.

Log in to comment