Why NVIDIA should be worried about Huawei's new AI architecture
What's the story
Huawei Technologies has introduced its new high-performance computing architecture, the Supernode 384. The company is actively promoting this innovative system among developers.
The Supernode 384 is positioned as a rival to NVIDIA's NVL72 system.
Huawei claims that its Supernode 384 can effectively tackle data center bottlenecks and meet future artificial intelligence (AI) training requirements with cutting-edge computing architectures.
Reason
Why is Supernode 384 necessary?
Zhang Dixuan, President of Huawei's Ascend computing business, emphasized the need for new computing architectures like Supernode 384.
He said traditional server architectures face a major bottleneck in training due to increasing parallel processing scales and cross-machine bandwidth.
The company hopes that its new architecture will help overcome these challenges and improve data center efficiency.
Tech specs
Supernode 384 powers CloudMatrix system
The Supernode architecture is the backbone of Huawei's CloudMatrix 384 system.
This advanced setup comprises a cluster of 384 Ascend AI processors, distributed across 12 computing cabinets and four bus cabinets.
Together, they deliver an impressive computing power of up to 300 petaflops and high-bandwidth memory capacity of up to 48TB.
Deployment
CloudMatrix 384: Huawei's largest AI training platform
Huawei has already deployed the CloudMatrix 384 system in its data centers across Anhui province, Inner Mongolia, and Guizhou province.
The company claims that this setup is the industry's largest AI training platform.
The deployment of such a powerful system underlines Huawei's commitment to pushing the boundaries of AI technology despite facing challenges from US tech control measures.