Vera Rubin 的 GPU 容量及吞吐量提升屬預測;Eos AI 工廠的示範則顯示,可在電網需求訊號出現時削減非緊急運算,同時維持高優先級推論。
What did Nvidia present on the opening day of its AI Infra Summit about transforming AI factories from systems optimized primarily for raw pNvidia’s AI Infra Summit focused on increasing useful AI output within fixed power budgets.
AI 提示
Create a landscape editorial hero image for this Studio Global article: What did Nvidia present on the opening day of its AI Infra Summit about transforming AI factories from systems optimized primarily for raw p. Article summary: Nvidia’s opening-day message was that an AI factory should be measured not only by peak compute, but by validated agentic tokens produced per megawatt—requiring integrated design across silicon, systems, networking, soft. 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
AI 基建而家愈嚟愈唔係「塞到幾多張 GPU 入機架」咁簡單,而係有冇足夠電力去推動佢哋。Nvidia 在 2026 年 AI Infra Summit 首日提出,評估 AI 工廠的關鍵產出指標,應該係 每兆瓦(MW)可驗證產出的代理式 AI Token 數量,而唔單止係峰值運算力。23
所謂 Token,可以粗略理解為 AI 模型處理或生成文字、程式碼等內容時的基本單位;對會自行拆解任務、反覆推理同調用工具的「代理式 AI」而言,持續、有效地產出 Token,先至係真正可交付的工作量。
Nvidia 超大規模及高效能運算副總裁 Ian Buck 在美國加州聖塔克拉拉的活動上,將焦點由單一 GPU 效能拉闊至整個系統:運算、網絡、推論軟件、散熱、電力控制,以至同電網協調,都要一齊設計,先可以喺電力成為硬約束時維持產能。123
在 Nvidia 的 Eos AI 工廠,Emerald AI 的 Conductor 平台與 Silicon Valley Power 合作參與商業化彈性負載計劃。Nvidia 指系統已回應超過 200 個公用事業公司發出的電力需求訊號;其中一次示範中,設施在不足 1 分鐘內把耗電量由 4 MW 降至 3 MW,同時透過降低可彈性處理工作的優先次序,讓較高優先級的推論繼續運作。23