🕷️
SetupAI
EN
🏠較早收集於 4h

英偉達測試 700W+ RTX 5090 Ti 原型

英偉達測試 700W+ RTX 5090 Ti 原型
PostLinkedIn
🏠閱讀原文： IT之家
#cuda-cores#high-tdp#blackwell-diertx-5090-ti

💡700W+ prototype reveals NVIDIA's power-hungry GPU future for AI training rigs.

⚡ 30-Second TL;DR

有什麼變化

22,848 個 CUDA 核心，比 RTX 5090 多 5%，但未達完整 GB202 的 24,576 個

為什麼重要

這凸顯 GPU 擴展的邊際效應遞減，促使 AI 基礎設施偏向效率而非原始功率。高 TDP 可能加重資料中心供電壓力，影響訓練叢集採購決策。

下一步行動

Benchmark your current RTX 5090 setup's power limits before planning Blackwell upgrades.

誰應關注：Developers & AI Engineers

關鍵要點

  • •22,848 個 CUDA 核心，比 RTX 5090 多 5%，但未達完整 GB202 的 24,576 個
  • •正常功耗 700-750W，解鎖原型超過 1000W
  • •需巨型散熱器，形容「雙手才能捧起」
  • •效能提升約 10%，理想條件下達 20%
  • •原型存在，但像過去 Titan 一樣可能永不發行

🧠 深度解析

背景與延伸：來自公開資料，非原文內容。引用 6 個來源。

🔑 增強重點摘要

  • •NVIDIA has been testing RTX 5090 Ti/Blackwell Titan prototype since mid-2025, with 22,848 CUDA cores (5% more than RTX 5090's 21,760), not using full GB202 die (24,064-24,576 cores).[1][2][3]
  • •Power draw is 700-750W under normal conditions, exceeding 1000W in unlocked prototypes, requiring massive cooling solutions.[1][2]
  • •Performance uplift is modest at 10-20% over RTX 5090, potentially due to higher clocks despite fewer cores than pro variants like RTX 6000.[1][2][3]
  • •No confirmed VRAM specs, but 32GB+ speculated; pricing could exceed $2500-3000, targeting enthusiasts and AI/gaming hybrid users.[1]
  • •Prototype existence noted, but commercial release uncertain, similar to unreleased Titan Ada; some reports question new RTX gaming GPUs in 2026.[1][2][6]
📊 競品分析▸ Show
FeatureNVIDIA RTX 5090 Ti/Blackwell TitanAMD Equivalent (Speculative)Intel Arc (Current Flagship)
CUDA Cores22,848N/A (RDNA 4 ~20,000+ est.)N/A (Xe-cores)
Power (TDP)700-750W~400-600W est.~300-450W
Performance vs 5090+10-20%N/A (no direct competitor)Significantly lower

🛠️ 技術深入

  • Uses modified GB202 chip, missing one GPC compared to full die in RTX 6000 (24,064 cores, 128MB L2 cache); RTX 5090 has 21,760 cores, 96MB L2.[1][3][4]
  • Blackwell architecture: 4th-gen RT cores, 5th-gen Tensor cores, Multi Frame Generation (MFG) as key new feature over Ada Lovelace.[2][5]
  • Expected 32GB GDDR7 on 512-bit bus (same as 5090), PCIe 5.0, DisplayPort 2.1b/HDMI 2.1a; potential dual-channel DDR5 IO limitations.[3][5]
  • Higher power enables modest gains via clock speeds, not core count; pro variants like RTX 6000 show limited gaming uplift despite more cores.[2]

🔮 前景展望AI analysis grounded in cited sources

High power demands (700W+) challenge PSU/cooling standards, may limit mainstream adoption; modest gains question value at premium pricing ($2500+), positioning as enthusiast/AI hybrid until RTX 60-series (2028); reflects NVIDIA's strategy amid no direct gaming competition.[1][2][6]

⏳ 時間線

2025-01
RTX 50-series (5070/5080/5090) announced at CES and released, based on Blackwell architecture with up to 21,760 cores and 32GB GDDR7.
2025-06
NVIDIA begins testing RTX 5090 Ti/Blackwell Titan prototypes.
2025-12
Driver 591.44 restores 32-bit PhysX support; reports of ROP anomalies in some RTX 50 cards resolved.
🏠閱讀原文： IT之家
📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

同主題

探索 #cuda-cores

同產品

更多關於 rtx-5090-ti

同來源

最新來自 IT之家

豆包輸入法 HarmonyOS 版開放搶先體驗

豆包輸入法 HarmonyOS 版開放搶先體驗

IT之家•Aug 22

AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： IT之家 ↗

這是摘要，不是原文。去看原站，或訂閱每週簡報。

每週 AI 簡報

每週一封，可隨時退訂。