华为汪涛：华为要打造AI算力底座，只做好一颗芯片远远不够

AI Digest

华为汪涛强调AI算力底座需系统工程，不仅提升芯片性能，更通过NPO光互联、超节点集群和CANN生态构建完整算力体系，为大模型训练提供稳定支撑。

Huawei's Wang Tao stresses AI infrastructure requires system engineering, enhancing chip performance while building comprehensive computing power through NPO optics, super nodes, and CANN ecosystem.

Key points

• 昇腾960DT单芯片算力倍增，支持4PFLOPS FP4和288GB HBM Ascend 960DT achieves doubled single-chip performance with 4PFLOPS FP4 and 288GB HBM
• 超节点集群通过统一内存编址实现多卡协同，MFU提升2.75倍 Super node clusters boost MFU 2.75x via unified memory addressing
• NPO光互联技术降低传输损耗，Hi-ONE光引擎实现7.2Tbps带宽 NPO optics reduce transmission loss with 7.2Tbps Hi-ONE engine
• CANN生态跨过拐点，月活开发者超5200人且61%为外部开发者 CANN ecosystem crosses inflection point with 5200+ active developers
• 华为定位为算力底座提供者，而非与模型厂商直接竞争 Huawei positions as compute foundation provider, not direct competitor

Takeaway: 华为通过系统工程构建AI算力底座，实现芯片-互联-生态的协同优化。 / Huawei builds AI compute foundation through chip-interconnect-ecosystem synergy.

Why it matters 揭示AI基础设施竞争本质，展现华为从芯片到生态的系统化布局策略。