Name: Nvidia Vera Rubin 深度解析 — 比 Blackwell 更高效 10 倍的接班人！ (Deconstructing Nvidia’s Vera Rubin — The Successor To Blackwell That’s 10x More Efficient)
Uploaded: 2026-03-02T22:10:57.000Z
Duration: 14 min

This is what Nvidia's Grace Blackwell looks like on the inside.

這就是 Nvidia 的 Grace Blackwell 內部樣子。

The 72 GPU system that's seen soaring sales from the likes of Microsoft, Google, Amazon and Meta.

這套擁有 72 個 GPU 的系統，銷售成績亮眼，像是 Microsoft、Google、Amazon 和 Meta 都買單。

程度副詞

But what we're really at Nvidia for is an exclusive first look at this.

但我們這次來 Nvidia，主要是為了搶先獨家看看這個。

Nvidia's next RackScale system for AI data centers has the AI world buzzing because Nvidia says it'll help solve the biggest bottleneck threatening AI buildouts today.

Nvidia 下一代用於 AI data centers 的 RackScale 系統，讓 AI 界議論紛紛，因為 Nvidia 表示它將有助於解決目前威脅 AI 建置的最大瓶頸。

It's actually going to be about 10 times more performant in terms of performance per watt compared to Blackwell.

相較於 Blackwell，它的 performance per watt 表現會強上大約 10 倍。

We're taking you on a deep dive inside Vera Rubin to find out.

我們要帶大家 deep dive Vera Rubin 內部，一探究竟。

Because this AI powerhouse takes a lot more companies than Nvidia and a lot more parts than these GPUs.

因為這個 AI powerhouse，需要的不只是 Nvidia，還有比這些 GPU 多得多的零件。

It also has the compute trays, the chassis, the side rails, the bus bars and 1.3 million components.

它還有 compute trays、chassis、side rails、bus bars，以及 130 萬個零件。

Over 80 different suppliers and across more than 20 countries.

來自超過 80 家不同的供應商，遍及 20 多個國家。

VR Rubin's in volume production today, with plans to ship later this year.

VR Rubin 現正量產中，預計今年稍晚出貨。

But from a memory shortage to a supply chain ridden with tariffs, what are the risks Nvidia faces to getting these two ton systems out the door?

但從記憶體短缺到充滿關稅的供應鏈，Nvidia 在將這兩噸重的系統出貨時會面臨哪些風險？

There is that risk that with all this AI capex, you basically have pulled forward or have a little bit too much than you need.

風險在於，由於所有這些 AI capex，你基本上是提前拉貨，或者擁有的比你需要的稍微多一點。

But I don't see that as being a risk this year.

And I think that's why Nvidia is so confident that their customers are saying we need more and if we had it, we could deploy it right now.

我認為這就是為什麼 Nvidia 對他們的客戶如此有信心，他們說我們需要更多，如果我們有，我們現在就可以部署。

From compute to connectivity to power, we take a look at the astounding ecosystem behind this single system and asked Nvidia how it built this complex web and whether competitors like AMD could ever catch up.

從運算到連接到電力，我們來看看這個單一系統背後令人驚嘆的生態系統，並詢問 Nvidia 如何建立這個複雜的網絡，以及像 AMD 這樣的競爭對手是否能追趕得上。

When Nvidia released its current Blackwell rack design in 2024, demand was insane.

Nvidia 在 2024 年發布其目前的 Blackwell rack 設計時，需求量大到不行。

Nvidia stock is up more than 100% since before it was announced.

Nvidia 的股價自宣佈前至今已上漲超過 100%。

Because it was the first full rack scale design that took this approach of disaggregating compute, the scale of networking as well as all the memory infrastructure.

因為它是第一個全機架規模的設計，採用了這種分解運算、網路規模以及所有記憶體基礎設施的方法。

But all of it was done to make it behave as a single gpu.

但這一切都是為了讓它能像單一 GPU 一樣運作。

The first generation of compute knew you'd buy a server, you'd buy a switch, you know you would, you would maybe buy a rack to cool if you got to a point where it needed that.

第一代的運算就知道你會買伺服器、買交換器，你知道的，你可能會買一個機架來散熱，如果到了需要的時候。

Got all these systems pulled together into a single rack, built for absolute greatest efficiency and greatest performance.

將所有這些系統整合到一個機架中，以實現最高的效率和最佳的效能。

And that's just not how Servers were historically built.

而這並非伺服器歷史上的建構方式。

Grace Blackwell has 72 GPUs, nearly 800 other chips, 1.2 million total components made at some 350 factories.

Grace Blackwell 擁有 72 個 GPU、近 800 個其他晶片，總計 120 萬個零件，由約 350 間工廠製造。

TSMC is developing the silicon and the core chips.

臺積電正在開發矽晶片和核心晶片。

We're working with Foxconn on a lot of the other rack assembly components.

我們正與鴻海合作開發許多其他的機架組裝零件。

Delta Electronics for a lot of liquid cooling elements.

Delta Electronics 提供許多液體冷卻元件。

The nozzles that connect on the end of the hoses are made by a number of different providers.

水管末端的噴嘴由多家不同供應商製造。

Whether that's Amphenol and getting the connectors and the copper, whether that's Vertiv and getting the cooling distribution systems.

無論是 Amphenol 提供連接器和銅，或是 Vertiv 提供冷卻分配系統。

With so many components, not every rack is touched by the exact same global suppliers, from China to Israel to Mexico, the us, Vietnam, Thailand.

有這麼多零件，並非每個機架都由完全相同的全球供應商提供，從中國到以色列、墨西哥、美國、越南、泰國。

So Nvidia created a standard reference design.

所以 Nvidia 創造了一個標準參考設計。

So the whole idea by creating a standard is it opens it up to the entire ecosystem.

透過創造標準，整個想法是它能開放給整個生態系。

There's PowerShells by Megmeet Lighton or Flex Power components by Infineon analog devices or STMicroelectronics.

有 Megmeet Lighton 的 PowerShells，或 Flex Power 的 Infineon、Analog Devices 或 STMicroelectronics 元件。

Chassis by Foxconn or Interplex Bizlink for bus bars, Pinda for rack manifolds, Coldplates by Auras ABC Boyd Cooler Master JPC to Rico Deal for power Whips.

鴻海或 Interplex Bizlink 的機箱，Bizlink 的匯流排，Pinda 的機架管路，Auras、ABC、Boyd、Cooler Master、JPC、Rico Deal 的冷卻板，以及電源線。

This supply chain has really become this core group of companies that are building this all together.

這個供應鏈已經真正成為一群核心公司，共同打造這一切。

And of course, they're all leaning in on the success of Nvidia.

當然，他們都倚賴 Nvidia 的成功。

We're still, you know, producing thousands of racks per week.

我們每週仍然生產數千個機架。

And so what I'll say is they're spoken for.

所以我想說的是，它們都已經被預訂了。

So now Nvidia is making the next big leap to Vera Rubin.

所以現在 Nvidia 正朝 Vera Rubin 邁出下一個大步。

With full production announced at CES in January.

並於一月在 CES 上宣佈全面量產。

This is a Rubin Pod 1152 GPUs in 16 racks, with about 100,000 more components than Grace Blackwell.

這是一個 Rubin Pod，內含 1152 個 GPU，裝在 16 個機架中，比 Grace Blackwell 多了大約 10 萬個元件。

V Vera Rubin uses about double the energy, but can deliver way more compute.

Vera Rubin 使用的能源大約是兩倍，但能提供更多運算能力。

The number of tokens it's generating is exponentially higher than last generation.

它產生的 token 數量比上一代呈指數級增長。

And so they've become very efficient for these, you know, these heavy workloads.

所以它們對於這些，你知道的，這些繁重的工作負載變得非常有效率。

To get there, Nvidia redesigned all six of the core chips.

為了達到這個目標，Nvidia 重新設計了所有六個核心晶片。

It delivers about 2x the performance per watt compared to our previous Gen Grace CPU.

它提供的 performance per watt 大約是我們上一代 Grace CPU 的兩倍。

This can actually deliver about 50 petaflops of AI performance, so about 2.5x performance.

這實際上可以提供大約 50 petaflops 的 AI 效能，所以大約是 2.5 倍的效能。

所以讓我們來 deep dive 看看一個機架。

It has 18 compute trays, each one starring two Vera Rubin superchips.

它有 18 個 compute trays，每個都搭載兩顆 Vera Rubin superchips。

Each has one Vera CPU, two Rubin GPUs and much more.

每個都有一個 Vera CPU、兩顆 Rubin GPU 等等。

So this particular superchip has 17,000 components.

這個超級晶片實際上包含 17,000 個元件。

And as you look at all these elements that have to connect to each other, a lot of these are just to help make sure that the power is is being modulated correctly.

當你看到所有這些需要互相連接的元件時，其中很多都是為了確保電源能被正確調節。

So we actually have memory units that can be applied and removed on the new Vera Rubin processor.

所以我們有可以應用和移除於新的 Vera Rubin 處理器上的記憶體單元。

Compared to the Grace Blackwell they were all sort of soldered in.

相較於 Grace Blackwell，它們都是焊死的。

But the real memory star lines the top and bottom of the Rubin GPU itself.

但真正的記憶體明星是 Rubin GPU 本身頂部和底部的記憶體。

Eight stacks of the latest high bandwidth memory HBM4 by the likes of SK, Hynix and Samsung.

八堆最新的高頻寬記憶體 HBM4，由 SK、Hynix 和 Samsung 等公司提供。

Is there any risk because it is in such short supply right now?

因為現在供應量非常短缺，有沒有風險？

The one thing that we've been very laser focused on is making sure our entire supply chain is aware and aware of not just sort of the design itself, but

我們一直非常專注於確保我們的整個供應鏈都瞭解，並且不僅僅是瞭解設計本身，

we're actually giving them very detailed forecasts.

我們其實會給他們非常詳細的預測。

We're aligning to make sure that everything we're shipping will be met by our supply chain.

我們會協調確保我們出貨的每一個東西，供應鏈都能配合。

Another risk factor, those 1.3 million components running workloads 24.

另一個風險因素，那 130 萬個元件在 24 小時運作時，

Some customers reported overheating issues in the early days of Blackwell deployments.

有些客戶回報 Blackwell 剛部署的初期有過熱問題。

Every component has to fit together perfectly.

And I think when we were doing some of those initial deployments there were things like you didn't get proper seating with the liquid cooling valve.

我覺得在做那些初期部署時，有些狀況像是液體冷卻閥沒有正確就位。

It was just actual implementation, user error.

這只是實際執行上的問題，是使用者錯誤。

And for the most part all the systems are fully deployed today, running hundreds of thousands of scales and without an issue.

大部分的系統都已全面部署，運行數十萬個 RackScale，而且沒有任何問題。

Vera Rubin compute trays have no hoses, cables or fans.

Vera Rubin compute trays 沒有水管、線纜或風扇。

It's Nvidia's first system that's 100% liquid cooled.

這是 Nvidia 首款 100% 液體冷卻的系統。

Will this mean that data centers have to have a really robust liquid cooling loop ready and in place before they can move from Blackwell to Vera Rubin?

這是否意味著資料中心在從 Blackwell 轉移到 Vera Rubin 之前，必須準備好一個非常穩健的液體冷卻迴路？

Essentially we are requiring as we build out the AI factories of the future that the vast majority will be leveraging this liquid cool based architecture.

基本上，我們在建立未來的 AI factories 時，會要求絕大多數都將採用這種基於液體冷卻的架構。

The superchip is covered with cold plates piping in water that keeps the components cool.

超級晶片覆蓋著冷卻板，透過水管將水導入，以保持元件涼爽。

This is the cool cold plates on the Vera CPU.

You also have the Rubin GPUs that are also underneath there.

你也有 Rubin GPU，它們也在下面。

Data centers currently consume a lot of water.

VR Rubin 會在這方面有所幫助嗎？

One of the counterintuitive savings is actually you use less water using liquid cooled systems.

一個反直覺的節省是，其實用液體冷卻系統會用更少的水。

The main reason is because you don't use the evaporative cooling technologies as much.

主要原因是因為你比較不需要用到蒸發冷卻技術。

And that allows you to consume much less water through a closed Loop system.

這讓你可以透過封閉迴路系統消耗更少的水。

A lot of the needs for cooling come because of just how much power is running through the system.

大部分的冷卻需求都來自於系統中運行的龐大電力。

It's about twice the amount of overall power, roughly about 220 kilowatts per rack.

大約是總電力需求的兩倍，每機架約 220 千瓦。

And so we had to redesign obviously the power delivery system.

所以我們當然得重新設計供電系統。

But most of all, efficiency is tied to how quickly each GPU can access the memory and processing on every other GPU simultaneously, which Nvidia solves with its own networking product, NVLink.

但最重要的是，效率取決於每個 GPU 能同時存取其他所有 GPU 的記憶體和處理器的速度，Nvidia 透過自家 NVLink 網路產品解決了這個問題。

This is the NVLink switch chip responsible for connecting all of these GPUs and CPUs together to have them behave as one.

這就是負責將所有 GPU 和 CPU 連接起來，讓它們像一個整體一樣運作的 NVLink 交換晶片。

So this actually doubled the line rate from again 1.8 terabytes per second to 3.6 terabytes per second.

所以這實際上將線速加倍了，從每秒 1.8 TB 提升到每秒 3.6 TB。

字幕列表影片播放

Nvidia Vera Rubin 深度解析 — 比 Blackwell 更高效 10 倍的接班人！ (Deconstructing Nvidia’s Vera Rubin — The Successor To Blackwell That’s 10x More Efficient)

sort

entire

incredible

process