Particle.news
Download on the App Store
14 ARTICLES
·
19h ago

Huawei Details Ascend AI Roadmap and ‘SuperPoD’ Mega-Clusters in Bid to Rival Nvidia

The rollout aligns with Beijing’s push to move AI computing onto domestic suppliers.

Image
Huawei has announced AI computing cluster for the first time.
Huawei's Ascend AI chips are becoming one of the company's most important products, especially given China's constrained access to U.S.-made AI chips.

Overview

  • Huawei introduced Atlas 950 and 960 SuperPoD systems that group 8,192 and up to 15,488 Ascend processors per node, with SuperClusters that the company says can scale from roughly 500,000 to over 1 million chips.
  • The company outlined Ascend 950, 960 and 970 chips through 2028, starting with Ascend 950PR in Q1 2026 for inference and 950DT in Q4 2026 for training, followed by Ascend 960 in 2027 and 970 in 2028.
  • Huawei said it has designed proprietary high-bandwidth memory, with HiBL 1.0 offering 128 GB at 1.6 TB/s for the 950PR and HiZQ 2.0 offering 144 GB at 4 TB/s for the 950DT.
  • Huawei claimed its new supernodes would be the world’s most powerful for years, while analysts note the strategy leans on clustering many domestic chips to offset lower per-chip performance versus Nvidia.
  • Reports say Chinese regulators have told tech firms to halt testing and purchases of certain Nvidia parts such as the RTX Pro 6000D, and Huawei also previewed Kunpeng 950/960 CPUs and a TaiShan 950 SuperPod for general-purpose computing.