Click the blue text
Follow us
NEWS TODAY
Huawei’s Ascend NPU cluster has broken through the training barrier for large models with nearly a trillion parameters, achieving stable training of a 718B parameter MoE model with over 6000 chips, increasing computing power utilization by 58.7%! Purely domestic hardware has smoothly overcome four major technical challenges including load balancing and communication walls, making hardcore black technology overshadow NVIDIA GPUs.
Huawei Ascend has successfully trained a large model with nearly a trillion parameters!
Previously, to train models close to a trillion parameters, everyone had to rely on NVIDIA GPUs. However, Huawei has achieved stable long-term training of a 718 billion parameter MoE (Mixture of Experts) model directly on its own Ascend NPU platform.
The latest technical report reveals that Huawei’s Pangu team (including the Noah’s Ark Lab and Huawei Cloud) has solved the “roadblocks” of load balancing, communication overhead, and inefficiency for running large models through multiple system-level optimizations on a cluster composed of over 6,000 Ascend NPUs.
With these innovative optimizations, the training speed has skyrocketed, directly supporting the development of top large models! From now on, large companies no longer need to rely on NVIDIA GPUs for training — the term “domestic” is becoming increasingly significant in the hardware for large models. This achievement not only refreshes the domestic computing power record but also signifies that our country has reached an internationally leading level in AI infrastructure.
IDC Observation
IDC Observation is a domestic vertical media and resource service platform focused on AIDC and intelligent computing fields. Relying on public accounts, brand mini-programs, and user-sharing ecosystems, it builds a full media matrix of “content + tools + community,” providing cutting-edge industry information, real-time query and sharing services for national data centers and computing power resources, accurately connecting data center operators, computing power demanders, and industry practitioners.
