Particle.news
Download on the App Store
19 ARTICLES
·
11h ago

Huawei Details AI Roadmap, In‑House HBM and ‘SuperPoD’ Megaclusters to Cut Reliance on Nvidia

Huawei touts scale‑first systems using SuperPoDs, proprietary HBM, UnifiedBus, reflecting Beijing’s push for domestic chips.

Image
Huawei has announced AI computing cluster for the first time.
SHANGHAI, CHINA - JUNE 29, 2020 - Huawei's largest flagship store in the world, Shanghai, China, June 29, 2020.- PHOTOGRAPH BY Costfoto / Barcroft Studios / Future Publishing (Photo credit should read Costfoto/Barcroft Media via Getty Images)

Overview

  • The company published a multi‑year Ascend roadmap: 950PR in Q1 2026, 950DT in Q4 2026, 960 in 2027, and 970 in 2028, alongside Kunpeng 950 and 960 CPUs for general computing.
  • Atlas 950 and 960 SuperPoDs are specified to group 8,192 and 15,488 Ascend chips per node, with SuperClusters scaling beyond 500,000 and over one million chips respectively.
  • Huawei disclosed proprietary HBM parts: HiBL 1.0 for Ascend 950PR (128 GB, 1.6 TB/s) and HiZQ2.0 for 950DT (144 GB, 4 TB/s), positioning them as cost‑effective versus current HBM3E/HBM4E.
  • UnifiedBus 2.0 was introduced as a new interconnect protocol to link SuperPoDs and SuperClusters, with Huawei claiming lower latency and fewer switches than RoCE‑only setups.
  • Analysts say Huawei’s strategy trades per‑chip performance for massive scale, as Chinese regulators reportedly discourage purchases of certain Nvidia China‑specific chips such as the RTX Pro 6000D.