What is Huawei's New AI Infrastructure Strategy?
Huawei’s David Wang, Deputy Chairman of the Board and Rotating Chairman, highlighted its commitment to building powerful AI infrastructure
TL;DR
- Huawei introduced the Atlas 960E SuperPoD, a liquid-cooled system powered by UnifiedBus and Hi-ONE, delivering 8 EFLOPS at FP8 and 16 EFLOPS at FP4 precision across 4,096 NPUs.
- The SuperCluster architecture can interconnect up to one million NPUs, and the TaiShan 950 SuperPoD has been upgraded to support up to 4,096 nodes with a 256TB unified memory pool.
- The OceanStor M900 storage system offers petabyte-scale KV cache and extends SSD lifespan by 16-fold, designed for agentic inference.
- Huawei is accelerating the release of its Ascend chip series, with Ascend 960DT and 960PR expected in 2027, and plans to launch Ascend 970 and 980 in 2028 and 2029.
- The company is committed to open source and open systems, with its Kunpeng ecosystem attracting millions of developers and its CANN architecture embracing community-driven development.
- Huawei plans to integrate Kirin and Ascend chips, enhance Pangu and third-party models, redefine HarmonyOS as an 'Agent OS,' and cultivate a diverse AI ecosystem for on-device computing platforms.