Name: Nvidia GPU、Google 和 Amazon AI 晶片大比拚！誰是 AI 霸主？ (How Nvidia GPUs Compare To Google’s And Amazon’s AI Chips)
Uploaded: 2025-12-02T00:01:33.000Z
Duration: 15 min 59 s

Nvidia graphics processing units like These latest Blackwell GPUs are inside server racks all over the world.

Nvidia 的圖形處理單元，像是這些最新的 Blackwell GPUs，就裝在世界各地的伺服器機架裡。

分詞構句

Nvidia has catapulted from gaming giant to the very core of generative AI, training the models, running the workloads and sending Nvidia's valuation soaring.

Nvidia 從遊戲巨頭一躍成為生成式AI的核心，負責模型訓練、運行工作負載，並讓Nvidia的估值飆升。

With 6 million Blackwell GPUs shipped over the last year, this connects all 72 GPUs, allowing the ACT as a single GPU to power the most advanced AI workloads.

過去一年出貨了六百萬個Blackwell GPUs，這將所有72個GPUs連接起來，讓ACT能像單一GPU一樣，運行最先進的AI工作負載。

GPUs are the general purpose workhorse stock of Nvidia and its top competitor amd.

GPUs是Nvidia和其主要競爭對手amd的通用工作馬股票。

But another big category of AI chip is gaining ground.

不過，另一大類的AI晶片也正在嶄露頭角。

The custom ASIC application specific integrated circuit now being designed by all the major hyperscalers.

也就是現在所有主要的Hyperscaler（超大規模雲端服務供應商）都在設計的客製化ASIC應用特定積體電路。

Google, Amazon, Meta, Microsoft, OpenAI alongside Broadcom.

包括Google、Amazon、Meta、Microsoft、OpenAI以及Broadcom。

These AI chips are smaller, cheaper, accessible and as the name suggests, specifically built for one purpose.

這些AI晶片體積較小、成本較低、更容易取得，而且正如其名，是專門為單一目的打造的。

But we see that growing even faster than the GPU market over the next few years.

但我們預期在未來幾年，這個市場的成長速度會比GPU市場更快。

AI chips also include FPGA's Field Programmable Gate arrays and an entire group of chips used to power edge AI on device instead of in the cloud.

AI晶片還包括FPGA的Field Programmable Gate arrays（現場可程式化邏輯閘陣列），以及一整群用於在裝置上運行邊緣AI（而非雲端）的晶片。

In short, there's a lot of AI chips and the space is only getting more crowded.

總之，AI晶片種類繁多，這個領域只會越來越擁擠。

So let's take a pause to break it all down.

所以我們來暫停一下，把這些東西好好拆解一下。

What's different good and bad about them?

Each and a brief look at all the major companies getting in on the AI chip making trend.

個別介紹，並簡要看看所有投入AI晶片製造趨勢的主要公司。

For a brief moment in October, Nvidia was the first company ever to reach a $5 trillion valuation thanks to its original gaming engine, the GPU becoming king of AI chips.

在十月的一個短暫時刻，Nvidia 成為史上第一家市值達到 5 兆美元的公司，這要歸功於它最初的遊戲引擎，也就是 GPU，成為了 AI 晶片的王者。

那麼，究竟是什麼讓晶片最適合 AI 呢？

GPU in this case was really purpose built to deliver parallel programming because when you think about rendering an image or in a scene, you need to calculate all those pixels at once.

GPU 在這個例子中，是為了實現平行程式設計而特別打造的，因為當你想像渲染圖像或場景時，你需要同時計算所有像素。

And so AI really lends itself to taking advantage of that capability.

所以 AI 真的非常適合利用這種能力。

To understand how GPUs became synonymous with AI, let's go back to 2012 to what many consider AI's big bang moment.

為了了解 GPU 如何與 AI 劃上等號，讓我們回到 2012 年，也就是許多人認為是 AI 的大爆炸時刻。

Alexnet it was a new, incredibly accurate neural network that obliterated the competition during a prominent image recognition contest.

Alexnet 是一個全新、極其精準的神經網路，在一個重要的圖像辨識競賽中徹底擊敗了所有競爭對手。

Turns out, the same parallel processing used by Nvidia GPUs to render lifelike graphics is also great for training neural networks, where a computer learns from data rather than relying on a programmer's code.

原來，Nvidia GPUs 用來渲染逼真圖形所使用的平行處理技術，對於訓練神經網路也非常有效，電腦能從數據中學習，而不是依賴程式設計師的程式碼。

The researchers took a GPU and said, I'm going to hack it to get it to expose the parallel computation capabilities to unlock that performance for the deep learning use case.

研究人員拿了一個 GPU，然後說，我要駭入它，讓它暴露其平行運算能力，以釋放深度學習應用場景的效能。

While AI workloads are usually accelerated on a GPU or custom asic, it still often needs a host central processing unit cpu, like the Grace in Nvidia's Grace Blackwell Server Rack system.

雖然 AI 工作負載通常在 GPU 或客製化 ASIC 上加速，但它仍然需要一個主機中央處理單元 CPU，就像 Nvidia Grace Blackwell 伺服器機架系統中的 Grace 一樣。

But while a CPU has a small number of powerful cores running sequential general purpose tasks, GPUs have thousands of smaller cores more narrowly focused on parallel math like matrix multiplication, used to process multidimensional data structures known as tensors.

但是，CPU 擁有少量強大的核心來執行序列化的通用任務，而 GPU 則有數千個較小的核心，更專注於平行數學運算，例如矩陣乘法，用於處理稱為張量的多維數據結構。

This ability to perform many operations simultaneously makes GPUs ideal for the two main phases of AI training and inference.

這種同時執行多項操作的能力，讓 GPU 成為 AI 訓練和推論這兩個主要階段的理想選擇。

Training is teaching the AI model to learn from patterns in large amounts of data, while inference uses the AI to make decisions based on new information.

訓練是教導 AI 模型從大量數據的模式中學習，而推論則利用 AI 根據新資訊做出決策。

The way AI shows up in everyday applications, whether it's the way you buy your coffee on your Starbucks app or how you interact with Salesforce for work, or speaking to us through our EarPods, that is being done through inference.

AI 出現在日常應用程式中的方式，無論是你在 Starbucks app 上買咖啡，或是在工作中與 Salesforce 互動，或是透過我們的 EarPods 與我們交談，這些都是透過推論來實現的。

Nvidia sells its GPUs directly to AI companies like in a recent deal to sell at least 4 million to OpenAI and to foreign governments including South Korea, Saudi Arabia and the UK.

Nvidia 直接將其 GPUs 出售給 AI 公司，例如最近以至少 400 萬套的交易賣給 OpenAI，以及出售給包括南韓、沙烏地阿拉伯和英國在內的各國政府。

It also sells GPUs to cloud providers like Amazon, Microsoft and Google, who go on to rent them out by the hour, minute or second.

它也將 GPUs 出售給 Amazon、Microsoft 和 Google 等雲端服務供應商，這些供應商接著以每小時、每分鐘或每秒為單位出租它們。

Today, systems that can handle AI workloads are in such high demand, One of these 72 Blackwell server racks sells for around $3 million, and

目前，能處理 AI 工作負載的系統需求量非常大，其中一個 Blackwell 伺服器機架就賣大約 300 萬美元，

Nvidia told us it's shipping 1,000 each week.

Nvidia 告訴我們，他們每週出貨 1,000 套。

And for Nvidia, they're looking to sell the entire system, not just the chip, because when you think at that system level, you get more efficiencies in terms of speed and power and performance.

對 Nvidia 來說，他們想賣的是整個系統，而不僅僅是晶片，因為從系統層級來看，速度、功耗和效能的效率會更高。

Of course, Nvidia isn't the only one making Rackscale GPU systems optimized for AI.

當然，Nvidia 並非唯一一家製造針對 AI 優化的機架級 GPU 系統的公司。

Its top competitor, Advanced Micro Devices AMD has seen major gains with its Instinct GPU line and major commitments from OpenAI and Oracle.

它的主要競爭對手 amd，其 Instinct GPU 系列已嶄露頭角，並獲得 OpenAI 和 Oracle 的大力支持。

A big difference in AMD GPUs is that they use a largely open source software EcoSystem, while Nvidia GPUs are tightly optimized around CUDA, Nvidia's proprietary software platform.

amd GPU 的一大特色是它們主要採用開源軟體生態系統，而 Nvidia GPU 則緊密圍繞著 Nvidia 的專有軟體平臺 CUDA 進行優化。

Nvidia says its next generation Rubin GPU will be in full production next year.

Nvidia 表示，其下一代 Rubin GPU 將於明年全面量產。

This is just an incredibly beautiful computer.

In the early boom days of large language models, compute intensive training has been key, a perfect job for GPUs, but

在大型語言模型早期蓬勃發展的日子裡，計算密集型訓練一直是關鍵，這對 GPUs 來說是完美的任務，但

as the models mature, more and more inference is needed.

隨著模型成熟，越來越多的推理需求出現。

Now post training techniques have been introduced to make these models very capable.

現在已經推出了後訓練技術，讓這些模型變得非常強大。

Now to extract the value you need to go to inference and so that's how people actually use AI, is through inference.

現在要提取價值，你就需要進行推理，這也是人們實際使用AI的方式，就是透過推理。

Inference can happen on less powerful chips programmed for more specific tasks.

推理可以在為更特定任務編程的、功能較弱的晶片上進行。

Enter custom asics, application specific integrated circuits and every major cloud provider designing their own.

這就引進了客製化 ASIC，也就是應用特定積體電路，而每個主要的雲端供應商都在設計自己的。

While a GPU is like a Swiss army knife, able to do many kinds of parallel math for different AI workloads, think of an ASIC like a single purpose tool, very efficient and fast, but

GPU 就像一把瑞士刀，能夠為不同的 AI 工作負載執行各種平行運算，但 ASIC 則像是單一用途的工具，非常有效率且快速，但是

hardwired to do the exact math for one type of job.

它是硬體鎖定，只能執行一種工作的確切運算。

They can be much, much more efficient in running those workloads, but

它們在運行這些工作負載時可以效率高得多，但是

when you specialize them, you can't change them once they're already carved into silicon.

當你專門化它們時，一旦它們被刻入矽晶片，你就無法改變它們了。

And so there's a trade off in terms of flexibility.

While Nvidia GPUs cost upwards of $40,000 and can be hard to get, startups rely on them because it's even more costly to make your own custom ASIC for AI, it is hugely expensive.

雖然 Nvidia 的 GPUs 要價數萬美元，而且還很難買到，但新創公司還是得依賴它們，因為自行打造客製化 AI ASIC 的成本更高，非常昂貴。

Costs minimum tens and often hundreds of millions of dollars.

最低成本也要數千萬，甚至數億美元。

But for the biggest cloud providers who can't afford them, custom Asics pay off because they're more power efficient and reduce reliance on Nvidia.

但對於那些買不起的超大型雲端服務供應商來說，客製化 ASIC 是划算的，因為它們更省電，而且能減少對 Nvidia 的依賴。

They want to bring cost of AI down, they want to have a little bit more control over the workloads that they build.

他們希望能降低 AI 的成本，並對他們建構的工作負載有更多一點的掌控權。

But at the same time they're going to continue to work very closely with Nvidia, with amd, because they also need the capacity.

但同時，他們也會繼續與 Nvidia、amd 密切合作，因為他們也需要產能。

Google was the first to make a custom ASIC for AI acceleration, coining the term Tensor Processing unit when its first ASIC came out in 2015.

Google 是第一家為 AI 加速打造客製化 ASIC 的公司，並在 2015 年推出首款 ASIC 時，創造了「Tensor Processing unit」這個詞。

The TPU also helped lead to the invention of the transformer at Google in 2017, the architecture powering almost all modern AI.

TPU 也促成了 2017 年 Google 發明瞭 Transformer，這是一種驅動幾乎所有現代 AI 的架構。

We got a tour at Google's chip lab in 2024.

我們在 2024 年參觀了 Google 的晶片實驗室。

There's actually four chips inside there.

Actually two of those are connected to a host and machine that has CPUs in it.

其實其中兩顆是連接到一個有 CPU 的主機和機器。

And then all these colorful cables are actually linking together all of the Trillium chips to work as one large supercomputer.

而這些五顏六色的線纜，其實是將所有的 Trillium 晶片連接起來，作為一個大型超級電腦運作。

A decade after the first tpu, Google just released Ironwood, its seventh generation and with it a big deal with Anthropic to train its LLM Claude on up to a million TPUs.

在第一顆 TPU 問世十年後，Google 剛推出了 Ironwood，這是他們的第七代產品，並與 Anthropic 達成了一項大協議，要在多達一百萬顆 TPU 上訓練其大型語言模型 Claude。

Some people even think that they're technically on par or even superior to Nvidia.

有些人甚至認為它們在技術上與 Nvidia 相當，甚至更勝一籌。

But traditionally Google has only used them for in house purposes.

但傳統上，Google 只將它們用於內部用途。

There's a lot of speculation that in the longer run Google might open up access to TPUs more broadly.

有許多猜測認為，長遠來看，Google 可能會更廣泛地開放 TPU 的使用權限。

Amazon Web Services was the next cloud provider to design its own AI chips after acquiring Israeli chip startup Annapurna Labs in 2015, where Ron Diamant has worked since day one.

在收購了以色列晶片新創公司 Annapurna Labs 後，Amazon Web Services 成為下一個設計自家 AI 晶片的雲端服務供應商，Ron Diamant 自第一天起就在那裡工作。

AWS announced Inferentia in 2018 and launched Trainium in 2022, now approaching its third generation.

AWS 在 2018 年推出了 Inferentia，並在 2022 年推出了 Trainium，現在已接近第三代。

Think of Trainium like a cluster of workshops with many small flexible tensor engines, while Google's TPU is like one big factory conveyor belt with one big rigid grid specialized for matrix math.

你可以把 Trainium 想像成一個由許多小型彈性張量引擎組成的車間集群，而 Google 的 TPU 則像一個大型工廠輸送帶，擁有一個大型、專門用於矩陣運算、結構剛性的網格。

The difference between the names is quite telling, but

名稱上的差異其實相當能說明問題，但是

over time we've seen that Trainium chips can serve both inference and training workloads quite well.

隨著時間推移，我們看到 Trainium 晶片能很好地同時處理推論和訓練工作負載。

On average we see that trainium provides between 30 and 40% better price performance compared to other hardware vendors in AWS.

平均來說，我們看到 Trainium 在 AWS 中，與其他硬體供應商相比，能提供 30% 到 40% 的價格效能優勢。

In October we went to Northern Indiana for the first on camera tour of Amazon's biggest AI data center, where Anthropic is training its models on half a million Trainium two chips behind me right now they are working on delivering these towers.

十月時，我們前往印第安納州北部，進行亞馬遜最大 AI 資料中心的首次鏡頭採訪，那裡 Anthropic 正在訓練其模型，使用我身後這半百萬顆 Trainium 2 晶片，他們目前正在交付這些機架。

Two of them together is going to make up one Ultra server.

其中兩個組合起來，將構成一臺 Ultra 伺服器。

And what's especially interesting is there are no Nvidia GPS GPUs in there, although AWS is filling its other data centers with lots of Nvidia GPUs to feed hungry AI customers like OpenAI.

特別有趣的是，裡面並沒有 Nvidia 的 GPUs，儘管 AWS 正在用大量的 Nvidia GPUs 填滿其其他資料中心，以滿足像 OpenAI 這樣對 AI 需求強勁的客戶。

The investment they're making in our platform is still significant and significantly more than you're seeing even in some of the upcoming ASIC deployments.

他們對我們平臺的投資仍然相當可觀，而且遠超乎你所見到的，甚至比一些即將部署的 ASIC 投資還要多。

And we can't talk about custom ASICs without mentioning Broadcom and its top competitor Marvell, which act as backend partners so clients don't have to hire full silicon teams to handle everything in house.

而談到客製化 ASIC，就不能不提 Broadcom 和它的主要競爭對手 Marvell，它們扮演著後端合作夥伴的角色，讓客戶不必聘請完整的矽團隊來處理所有內部事務。

All of the big hyperscalers actually that have ASIC programs partner with at least one chip design company, Broadcom being the most important, which helps provide the IP and the know how and the networking.

實際上，所有擁有 ASIC 計畫的大型雲端服務供應商都會與至少一家晶片設計公司合作，其中 Broadcom 最為重要，它提供 IP、專業知識和網路技術。

And so you've seen Broadcom in particular be one of the biggest beneficiaries of the AI boom.

是以，你可以看到 Broadcom 尤其成為 AI 繁榮的最大受益者之一。

Broadcom helped build Google's TPUs, Meta's training and inference accelerator launched in 2023 and now a huge new deal to help OpenAI build its own custom ASICs starting in 2026.

Broadcom 協助建置了 Google 的 TPU、Meta 在 2023 年推出的訓練和推論加速器，現在還有一個重大的新協議，將從 2026 年開始協助 OpenAI 建置自己的客製化 ASIC。

We see Broadcom winning 70, maybe even 80% of this market.

我們看到 Broadcom 在這個市場贏得了 70%，甚至可能高達 80% 的市佔。

We see this market accelerating at a mid double digit CAGR over the next five years.

我們預期未來五年，這個市場的複合年增長率將會是雙位數中期。

Microsoft is also getting into the ASIC game, hoping to primarily use its own Maya chips in its Azure data centers, although little has been revealed since plans were announced in 2023 and its next Maya chip is facing delays.

Microsoft 也加入了 ASIC 的戰局，希望能主要在其 Azure 資料中心使用自家的 Maya 晶片，儘管自 2023 年公佈計畫以來，相關細節不多，且下一代 Maya 晶片面臨延遲。

Intel has its own Goudy line of custom Asics for AI.

Intel 則有自己的 Goudy 系列客製化 AI ASIC。

Tesla has announced an ASIC And Qualcomm is breaking into data center chips with the AI 200.

Tesla 已經宣佈了 ASIC，而 Qualcomm 則以 AI 200 晶片進軍資料中心晶片市場。

Of course, there's also a slew of startups going all in on custom AI chips too, like Cerebras making huge full wafer AI chips and Grok with inference focused language processing units.

當然，還有許多新創公司全力投入客製化 AI 晶片，例如 Cerebras 打造大型全晶圓 AI 晶片，Grok 則專注於推論語言處理單元。

The final big category of AI chips are those made for running edge AI on device instead of in the cloud.

最後一類大型 AI 晶片，是那些用於在裝置端運行邊緣 AI，而非在雲端運行的晶片。

You might want to be able to run an AI model locally on your personal device, even a smartphone.

你可能會想在個人裝置上，甚至手機上在地運行 AI 模型。

And so there may be as part of the processor in that chip a module that has AI processing capability, so you don't have to have communication all the way back to a data center and you can preserve privacy of your data on your phone.

是以，該晶片處理器中可能會有一個模組，具備 AI 處理能力，這樣你就不必一路連線回資料中心，也能保護手機上的資料隱私。

In npu, Neural processing unit is a major edge AI chip, a dedicated AI accelerator that's integrated into a phone or laptop's primary chip, what's called an SoC.

NPU，神經處理單元，是一種主要的邊緣 AI 晶片，它是整合到手機或筆記型電腦主要晶片（稱為 SoC）中的專用 AI 加速器。

A system on a chip which is kind of a chip that has like many different modules in it that do different aspects of the computation.

系統單晶片（SoC）就像一個晶片，裡面有很多不同的模組，負責不同的運算。

Because obviously in a phone you need to have a very compressed system that is able to do lots of things.

因為很明顯地，手機需要一個非常精簡的系統，而且要能做很多事情。

And so it's not kind of a separately packaged AI chip.

所以它不是那種獨立封裝的 AI 晶片。

They take up less silicon and therefore they cost less, often substantially less than a big data center chip would.

它們佔用的矽空間比較小，是以成本也比較低，通常比大型資料中心晶片的成本低很多。

NPU's enabling AI and PCs are primarily made by Qualcomm, intel and AMD for MacBooks.

啟用 PC 上 AI 的 NPU 主要由 Qualcomm、intel 和 AMD 為 MacBooks 生產。

Apple's in house M series chips include a dedicated neural engine too, although Apple doesn't use the term NPU.

Apple 自家的 M 系列晶片也內建了專用的神經引擎，不過 Apple 並沒有使用 NPU 這個詞。

Similar neural accelerators, tiny processors dedicated to AI math, are now built into the latest iPhone A series chips too, which we got to see in September.

類似的神經加速器，也就是專門處理 AI 數學運算的小型處理器，現在也內建在最新的 iPhone A 系列晶片裡了，我們在九月時有看到。

We know that when we can do things on device, we are able to manage people's privacy in the best way.

我們知道，當我們能在裝置端處理事情時，就能以最好的方式管理人們的隱私。

The other thing about it is it is efficient for us, it is responsive.

另外，對我們來說，它也很有效率，而且反應快速。

字幕列表影片播放

Nvidia GPU、Google 和 Amazon AI 晶片大比拚！誰是 AI 霸主？ (How Nvidia GPUs Compare To Google’s And Amazon’s AI Chips)

specific

process

dedicated

acquire