What did Inspur Information announce about the MetaBrain SD200 Ultra at AICC2026, including its intended workloads, 128-chip and memory confAI-generated illustration; not a photograph of the MetaBrain SD200 Ultra.
AI 提示詞
Create a landscape editorial hero image for this Studio Global article: What did Inspur Information announce about the MetaBrain SD200 Ultra at AICC2026, including its intended workloads, 128-chip and memory conf. Article summary: Inspur Information announced the MetaBrain SD200 Ultra at AICC2026 as a tightly coupled supernode for large frontier models and latency-sensitive AI-agent workloads. It also introduced the MetaBrain HC2000 as a separate,. 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
浪潮信息在 2026 年人工智慧運算大會(AICC2026)發表元腦 SD200 Ultra,定位為運行超大型模型、處理對延遲敏感的 AI 代理(Agent)任務的超節點。同場亮相的 HC2000 則瞄準另一種需求:在相同投資下產出更多推論詞元(token)。兩款產品最醒目的效能數字都來自廠商,採購時不宜直接當成可重現的實測結果。31820
128 顆晶片,能承載多大的模型?
SD200 Ultra 採用 3D Hyper Mesh 架構,緊密連接 128 顆本土 AI 加速晶片,配置 8 TB 統一編址加速器記憶體及 64 TB 系統記憶體。浪潮信息稱,單機可運行 2.8 兆參數的 Kimi K3,並支援高達 10 兆參數的前沿模型。後一數字說的是宣稱可承載的模型規模,不能據此推斷該規模下的推論速度。317
針對 Kimi K3,浪潮信息公布的數字是每生成一個詞元低於 5.85 毫秒,換算約為單一用戶每秒 170 個詞元,並稱速度達「業界平均」的五倍。但現有資料不足以確認比較雙方採用相同測試條件;逐詞元生成時間,也不同於首個詞元出現前的等待時間,更不是整段回答完成所需的時間。13