Name: Nvidia Kyber / Rubin Ultra 登場！單機櫃塞入 576 顆 GPU，搭配 Vera CPU！ (Nvidia Kyber / Rubin Ultra and NVL576 on display — 576 GPUs in a single rack with Vera CPUs)
Uploaded: 2026-02-07T00:16:12.000Z
Duration: 5 min 43 s

At the heart of Nvidia's latest innovation is the upcoming NVL 576 Rack, a single cabinet housing 576 Rubin Ultra GPUs paired with variseries CPUs.

Nvidia 最新創新核心是即將推出的 NVL 576 Rack，這是一個單一機櫃，可容納 576 個 Rubin Ultra GPUs 和 variseries CPUs。

All of this is built on the new Kyber rack architecture, representing a fundamentally new scale of computing density, power, cooling and system design.

這一切都建構在全新的 Kyber 機架架構上，代表著計算密度、功率、散熱和系統設計的全新規模。

This isn't just faster hardware, it's the kind of machine data centers build when they're thinking AI factory, not just AI server.

這不只是更快的硬體，而是資料中心在思考 AI 工廠時，而不僅是 AI 伺服器時會打造的機器。

The Kyber rack generation replaces the previous Oberon style racks, moving away from conventional horizontally mounted trays.

Kyber 機架世代取代了先前的 Oberon 風格機架，擺脫了傳統的水準安裝託盤。

Instead, Kyber uses vertical blade modules and tight networking midplanes, optimizing space and performance.

相反地，Kyber 使用垂直刀片模組和緊密網路中介平面，優化空間和效能。

Each of the four pods inside the rack takes in 18 blades, with each blade loading up to eight Rubin Ultra GPUs and accompanying VARICPU's.

機架內的四個儲存單元各包含 18 個刀片，每個刀片可載入高達八個 Rubin Ultra GPUs 和搭配的 VARICPU's。

This configuration adds up to an astonishing 576 GPUs in a single rack, providing unprecedented computing power.

此配置加起來，單一機架就擁有驚人的 576 個 GPU，提供前所未有的計算能力。

High speed NVLink and custom midplanes enable each GPU to communicate with many others with very low latency.

高速 NVLink 和客製化中介平面讓每個 GPU 都能以極低的延遲與許多其他 GPU 溝通。

This makes the rack behave more like one huge compute plane than dozens of separate boxes, a truly game changing approach.

這讓機架更像是一個巨大的計算平面，而不是數十個獨立的盒子，這是一個真正改變遊戲規則的方法。

With 576 high end GPUs in one rack, you no longer need dozens of racks to build a large model training or inference system.

在一個機架中擁有 576 個頂級 GPU，您就不再需要數十個機架來建構大型模型訓練或推理系統。

The NVL576 aims for exaflop class compute in one footprint, a massive scale in a single compact unit.

NVL576 旨在單一空間內達到 Exaflop 等級的計算能力，這是一個單一緊湊單元中的龐大規模。

This rack is planned to draw up to 600 kilowatts, far above current rack power budgets, necessitating a power and cooling rethink.

此機架規劃的功率最高可達 600 千瓦，遠高於目前的機架功率預算，需要重新思考功率和散熱。

Nvidia is shifting to 800 VDC power delivery with reducing copper volume and improving efficiency a critical advancement.

Nvidia 正轉向 800 VDC 供電，減少銅的使用量並提高效率是一項關鍵的進展。

Full liquid cooling solutions are integrated with the entire system, designed to handle the extreme thermal loads.

全面的液體散熱解決方案已整合到整個系統中，旨在處理極端的熱負荷。

Vera CPUs are tightly paired with Rubin GPUs, meaning the rack is built not just for throughput but for reasoning class tasks and agentic AI.

Vera CPUs 與 Rubin GPUs 緊密配對，這意味著該機架不僅用於吞吐量，還用於推理類任務和代理式 AI。

We're not just ramping up raw flops, we're building for next gen AI workloads including large context inference.

我們不僅在提升原始浮點運算能力，還在為包括大型上下文推理在內的下一代 AI 工作負載構建。

With 576 GPUs all in one rack, latency and bandwidth between GPUs are paramount, requiring a networking and interconnect redesign.

在一個機架中擁有 576 個 GPU，GPU 之間的延遲和頻寬至關重要，需要重新設計網路和互連。

NVLink switchblades, high bandwidth midplanes and optimized intra rack communication are critical to prevent bottlenecks for large language models.

NVLink switchblades、高頻寬中介平面和優化的機架內通訊對於防止大型語言模型的瓶頸至關重要。

Multimodal AI or Agentix systems, the NBL576 rack offers unprecedented capability.

無論是多模態 AI 還是 Agentix 系統，NBL576 機架都提供了前所未有的能力。

Teams can train or serve models at a scale they use to build across many racks or even data centers, all within a single unit.

團隊可以在單一單元內，以他們過去需要跨越多個機架甚至資料中心才能達到的規模來訓練或部署模型。

Inference latency drops, context windows grow, and model size and complexity can increase without splitting across many systems.

推論延遲降低，上下文視窗擴大，模型大小和複雜度也能增加，而無需分割到多個系統。

However, this advanced technology comes with major infrastructure demands requiring significant upgrades up to 600kW means facilities need upgraded electrical supply, high voltage distribution lines and special bus bars.

不過，這項先進技術伴隨著巨大的基礎設施需求，需要顯著升級，高達 600kW 表示設施需要升級電力供應、高壓配電線和專用匯流排。

At that density, water cooling or immersion cooling is no longer optional.

在這樣的密度下，水冷或浸沒式冷卻不再是可選的。

Traditional air cooled systems won't suffice.

傳統氣冷系統將無法滿足需求。

The rack footprint isn't just the cabinet.

機架的佔地空間不只是機櫃本身。

Sidecars for power conversion, chillers and monitoring gear must be considered.

電源轉換、冷卻器和監控設備的側掛單元也必須考慮進去。

Deploying such racks takes planning, capital, skilled operations and time to integrate, representing a big leap for many organizations.

部署這類機架需要規劃、資金、熟練的操作以及整合時間，對許多組織來說是一大躍進。

Enterprises aiming to build large scale AI factories now have a target rack of compute, consolidating into fewer ultra dense racks.

目標是建立大型 AI 工廠的企業，現在有了目標機架的計算能力，整合到更少的超高密度機架中。

Developers of generative AI, multimodal reasoning systems and real time inference pipelines gain a new lever.

開發生成式 AI、多模態推理系統和即時推論管線的開發者，獲得了一個新的槓桿。

More local compute per rack, fewer network hops and more scale are now achievable.

每個機架更多的在地計算、更少的網路跳數以及更大的規模現在都變得可行。

Accelerating innovation Data center design becomes even more critical with power, cooling, networking and maintenance all evolving.

加速創新 資料中心的設計變得更加關鍵，因為功率、散熱、網路和維護都在不斷演進。

For chip makers and infrastructure vendors, this rack signals a shift.

對於晶片製造商和基礎設施供應商來說，這個機架標誌著一個轉變。

The bottleneck is no longer just more gpu.

It's how you distribute power, how you cool and how you interconnect.

而是你如何分配功率、如何散熱以及如何互連。

If you're building something that only needs a few GPUs or moderate inference loads, this type of rack may be overkill.

如果你正在建置只需要少量 GPU 或中度推理負載的系統，這種機架可能有點過頭了。

But if you're planning AI at a large scale, you should consider what it means when one rack can do what formerly required dozens.

但如果你正在規劃大規模的 AI，你應該考慮一下，一個機架能做到過去需要數十個機架才能完成的工作，這意味著什麼。

GPU capability matters, but so does power architecture, interconnect, cooling and software model design needs to adapt.

GPU 的能力很重要，但功率架構、互連、散熱和軟體模型的設計也需要隨之調整。

With such dense racks, you can push models that need high memory bandwidth and large context windows.

透過如此密集的機架，你可以推動那些需要高記憶體頻寬和大型上下文視窗的模型。

Planning ahead is crucial if you're in procurement, operations or strategy.

如果你是採購、營運或策略人員，提前規劃非常重要。

Power upgrades, cooling systems and rack level design must be aligned with future racks like the NVL576.

功率升級、散熱系統和機架等級的設計，都必須與 NVL576 這類未來的機架相符。

In summary, the Kyber, Rubin Ultra and NBL576 rack is a preview of the next frontier of AI infrastructure.

總結來說，Kyber、Rubin Ultra 和 NBL576 機架預示了 AI 基礎架構的下一個前沿。

It takes the concept of cluster and shrinks it into a single rack with ultra dense computer, bespoke power cooling and high bandwidth interconnect.

它將叢集的概念濃縮到單一機架中，具備超高計算密度、專屬的功率散熱和高頻寬互連。

While planned for 2027, it sets a target for what AI factory hardware will look like, signaling how infrastructure and compute scale will evolve together.

雖然預計在 2027 年推出，但它為 AI 工廠硬體未來的樣貌設定了目標，預示著基礎架構和計算規模將如何一同演進。

Nvidia Kyber / Rubin Ultra 登場！單機櫃塞入 576 顆 GPU，搭配 Vera CPU！ (Nvidia Kyber / Rubin Ultra and NVL576 on display — 576 GPUs in a single rack with Vera CPUs)