五項部署以 SuperPoD 和 SuperCluster 為編程尺度,嘗試處理長時間運作的 AI 代理在回應速度、跨節點傳輸和記憶體方面的難題;實際成效仍須按工作負載驗證。[2][6]
At HUAWEI CONNECT 2026, why did Huawei’s Zhu Zhaosheng say Ascend had passed an ecosystem inflection point, and how do his five proposed shiAI-generated editorial illustration; not a photograph of Huawei hardware or the HUAWEI CONNECT event.
AI 提示
Create a landscape editorial hero image for this Studio Global article: At HUAWEI CONNECT 2026, why did Huawei’s Zhu Zhaosheng say Ascend had passed an ecosystem inflection point, and how do his five proposed shi. Article summary: Zhu’s “inflection point” claim was that Ascend was becoming more than a hardware platform: Huawei said its open software stack, developer community, and shared compute were now substantial enough for others to build on i. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
openai.com
一個 AI 代理如果要分多步完成長任務,難題唔係單靠一粒晶片「計得快」就解決到:任務暫停後要接得返上次進度,重新開始回應時要盡快輸出第一個 token;當工作分散到更多 NPU(AI 運算處理器),節點之間的資料傳輸和記憶體需求亦會愈來愈重要。26
在 HUAWEI CONNECT 2026,華為計算業務策略負責人朱照生稱,昇騰(Ascend)已跨過「生態轉捩點」。他的意思並非單指硬件有所進步,而是認為周邊的軟件、開發者社群和共享算力,已開始具備讓更多夥伴在昇騰之上開發的條件。這是華為對自身生態的判斷,並非長時間運作的 AI 代理效能已獲獨立證實。221
五項部署,分別想解決甚麼?
將跨節點資源組合起來。 華為提出以 API 組合 SuperPoD 和 SuperCluster 內不同節點的運算、通訊及記憶體資源,讓開發者較容易表達跨節點工作。SuperPoD 可理解為互連的大型運算系統,SuperCluster 則把規模推至叢集層面。目標是應付代理工作擴大時的資料傳輸和記憶體壓力,並不代表互連成本會就此消失。24