英偉達Vera Rubin

· 頻道:NVIDIA · 閱讀 19966

英偉達Vera Rubin

1. 英偉達在MLPerf Training v6.1中首次讓Vera Rubin NVL72機架級平臺登場,官方稱其訓練性能最高可達Blackwell Ultra的2倍。這輪基準的看點是把「訓練要多久」擺到檯面:AI智能體承擔的工作流越來越複雜,模型和智能體本身的迭代速度都取決於單輪訓練時長;NVL72走的是整機架互聯、以規模換吞吐的路線,因此形成與上代Blackwell Ultra的代際對照。我的看法是,2倍是官方基準條件下的數字,真實業務能不能達到這個幅度,要看互聯、散熱與軟件棧適配,採購前用自家模型規模複測一次更穩。

2. 原文內容:As AI agents tackle increasingly complex workflows, reducing time-to-train is critical for rapidly iterating on both models and the agents they power.⋯⋯(後半略)

3. NVIDIA's Vera Rubin NVL72 rack-scale system posted up to 2x the training throughput of Blackwell Ultra in MLPerf Training v6.1, the first benchmark round to include the new platform. The company argues that cutting time-to-train is now the main lever for iterating on models and the agents built on them. Full results and per-benchmark numbers are in NVIDIA's technical blog.

該內容經過二次創作,已經去除違法違規內容。

#英偉達 #AI訓練 #MLPerf #算力硬件

相關文章