Name: 與 Rebellions 公司首席執行官探討人工智能芯片的未來 (The future of AI chips with the CEO of Rebellions)
Uploaded: 2025-07-25T18:36:57.000Z
Duration: 29 min 9 s

Welcome to Beyond the Valley, a CNBC Tech podcast with Arjun Karpal and Tom Chitty.

歡迎收聽 CNBC 科技播客阿瓊-卡帕爾（Arjun Karpal）和湯姆-奇蒂（Tom Chitty）的《超越山谷》（Beyond the Valley）節目。

分詞構句

Chips are an integral part of the AI boom, powering the development of more advanced systems.

芯片是人工智能蓬勃發展不可或缺的一部分，為更先進系統的開發提供了動力。

In relation to semiconductors and AI, the name on everyone's lips is NVIDIA, which for a while has been seen as the only game in town.

在半導體和人工智能領域，英偉達（NVIDIA）是人人皆知的名字。

But there are a number of startups looking to challenge NVIDIA's dominance.

但是，也有一些初創公司希望挑戰英偉達的主導地位。

One of those is Rebellions, which last year became South Korea's first AI chip unicorn, or company valued over a billion dollars, after it merged with another semiconductor startup called Sapion.

Rebellions 公司就是其中之一，去年它與另一家名為 Sapion 的半導體初創公司合併後，成為韓國首家人工智能芯片獨角獸公司，即估值超過 10 億美元的公司。

Rebellions has some hefty backers, including SK Group and Saudi Aramco's venture capital arm.

Rebellions 擁有一些強大的支持者，包括 SK 集團和沙特阿美石油公司的風險投資部門。

My guest today is the CEO of Rebellions, and he's worked in companies from Intel to SpaceX.

我今天的嘉賓是 Rebellions 公司的首席執行官，他曾在英特爾和 SpaceX 等公司任職。

Sunghyun Park, thanks for joining me on this special episode of Beyond the Valley from the Mobile World Congress in Barcelona.

成賢（Sunghyun Park），感謝您在巴塞羅那世界移動通信大會期間收看本期《超越山谷》特別節目。

Let's just start off by talking about what does Rebellions do?

首先，讓我們來談談 Rebellions 是做什麼的？

First of all, thank you for inviting me here.

我是《山谷之外》的忠實粉絲。

It's an honor to be part of this program.

Yes, so as you mentioned, Rebellions was founded in September 2020.

是的，正如你提到的，Rebellions 成立於 2020 年 9 月。

And in Series A, we are backed by Korea Telecom.

在 A 輪融資中，我們得到了韓國電信公司的支持。

And right now we got merged with SK Sapion last December.

去年 12 月，我們與 SK Sapion 合併。

Right now we are backed by number one and two Telco in Korea.

目前，我們得到了韓國第一和第二大電信公司的支持。

That's why the MWC is our best place, you know, promotion our product and technology.

這就是為什麼 MWC 是我們推廣產品和技術的最佳場所。

What are your key products at the moment?

It's something like, you know, somebody's not very familiar with the AI chip.

你知道，有人對人工智能芯片不是很熟悉。

It's something like NVIDIA products, so junior NVIDIA or Korean version of NVIDIA.

它類似於英偉達（NVIDIA）的產品，所以是初級英偉達（NVIDIA）或韓版英偉達（NVIDIA）。

But in terms of products, we are focused on inference.

但就產品而言，我們的重點是推理。

But let me talk about the functional diversities.

不過，還是讓我來談談功能的多樣性吧。

NVIDIA's GPGPU, general purpose GPU offers much higher functional diversities.

英偉達™（NVIDIA®）公司的 GPGPU（通用 GPU）提供了更高的功能多樣性。

They can do any kinds of parallel computation, including training, inference, Bitcoin mining, or graphics rendering, any kinds of high-performance computing.

它們可以進行任何類型的並行計算，包括訓練、推理、比特幣挖礦或圖形渲染等任何類型的高性能計算。

However, when it comes to inference only, I think our product and technology offer a higher energy efficiency and performance, especially in latency.

不過，如果只涉及推理，我認為我們的產品和技術能提供更高的能效和性能，尤其是在延遲方面。

此外，我們還是一家韓國初創企業。

As you mentioned, we are Korea's only deep tech unicorn so far.

正如您提到的，我們是韓國迄今為止唯一一家深度科技獨角獸企業。

We are a fully-elaborated Korean semi-con ecosystem.

我們是一個完善的韓國半康生態系統。

That's a little bit different from other AI chip startups and NVIDIA.

這與其他人工智能芯片初創公司和英偉達有點不同。

And I do want to dig into, because what you're talking about is two different things.

我確實想深入瞭解一下，因為你們所說的是兩碼事。

And just if our listeners don't know, you have the training, right?

如果我們的聽眾不知道，你接受過培訓，對嗎？

We've spoken so much about training these huge, large language models, these huge AI foundational models.

我們已經說了很多關於訓練這些龐大的大型語言模型，這些龐大的人工智能基礎模型的事情。

And NVIDIA has been very successful in this particular market.

英偉達在這一特定市場上取得了巨大成功。

Now, a lot of companies are talking about inferencing.

現在，很多公司都在談論推理。

Now, this is really maybe the best way to think about it.

現在，這也許是最好的思考方式。

And you can correct me if I'm wrong, is the actual real application of the AI, right?

如果我說錯了，你可以糾正我，這就是人工智能的實際應用，對嗎？

When we talk about inferencing, why does the architecture, the chip architecture need to be different?

當我們談論推理時，為什麼架構、芯片架構需要不同？

Let me back to your very first question here.

For those who are not very familiar with the training and inference.

對於那些不太熟悉訓練和推理的人來說。

Training is kind of research and development to develop the AI models.

訓練是開發人工智能模型的一種研發工作。

On the other hand, inference is where actual business began.

另一方面，推論是實際業務的起點。

You know, from compute workload wise, it's something like feedback, peer through something that's too much complicated.

要知道，從計算工作量的角度來看，這就像是反饋，通過過於複雜的東西進行同行交流。

But from enterprise point of view, that's the R&D and the business.

但從企業的角度來看，這就是研發和業務。

That's why we Rebellion's focus is on the inference.

這就是為什麼我們《反叛》的重點是推論。

不過，讓我再回到你的問題上來。

Inference doesn't require any weight update during all the process workload, meaning that it doesn't need to be connected to multiple GPUs and not hundreds of the GPU or something like that.

推理不需要在所有進程工作負載期間更新權重，這意味著它不需要連接多個 GPU，也不需要數百個 GPU 或類似的東西。

A couple of GPUs or four to eight, because even with 175 billion models, just H100, for example, with the quantizer model, you know, a couple of H100 from NVIDIA is good enough.

幾個或四到八個 GPU，因為即使是 1,750 億個模型，僅以 H100 為例，使用量化模型，你知道，英偉達的幾個 H100 就足夠了。

So interconnects and other software flexibility and training require much higher flexibility, even in hardware and the software.

是以，互聯和其他軟件的靈活性和培訓需要更高的靈活性，甚至是硬件和軟件的靈活性。

Meaning from inference point of view, that's all overhead.

從推論的角度來看，這都是開銷。

We don't need that much flexibility and luxury.

我們不需要那麼大的靈活性和奢侈性。

So it's exactly what deep sick was doing.

所以，這正是 "深病 "在做的事情。

Deep sick developed their model using NVIDIA GPU.

深海病蟲利用英偉達™（NVIDIA®）圖形處理器開發了他們的模型。

On the other hand, as you know, the deep sick is top ranked in App Store, Apple.

另一方面，如你所知，"深度病症 "在蘋果應用商店排名第一。

There are a lot of traffic in inference, but they're doing the inference using Huawei's inference chip.

推理的流量很大，但他們使用華為的推理芯片進行推理。

That's exactly what approach and technically a little bit different.  So fascinating.

這正是方法和技術上的一點不同。  太迷人了

So you're able to do inferencing with fewer chips, is what you're saying?

你的意思是，你可以用更少的芯片進行推理？

Yes, without much fewer interconnects, much fewer SRAM.

是的，沒有更少的互連，更少的 SRAM。

But we still need some flexibility to support diversity of all AI inference demand.

但我們仍然需要一些靈活性，以支持所有人工智能推理需求的多樣性。

So our key technology, which is differentiated from other AI chip companies out there, is HBM centric and chip-led centric architecture.

是以，我們與其他人工智能芯片公司不同的關鍵技術是以 HBM 為中心、以芯片為主導的架構。

So NVIDIA is really doing great in the high bandwidth memory architecture.

是以，英偉達在高帶寬內存架構方面確實做得很出色。

But most of the United States AI chip startup, you know, very fancy, really good startup out there is, for example, Samba, Nova, Grow, TensorFlow, really good, really good talent.

但美國大多數人工智能芯片初創公司，你知道的，非常花哨、非常優秀的初創公司都在那裡，比如，Samba、Nova、Grow、TensorFlow，真的非常優秀，人才濟濟。

But they're not that much optimized for high bandwidth memory.

但它們對高帶寬內存的優化程度並不高。

But we are focused on high bandwidth memory and chip-led.

但我們專注於高帶寬內存和芯片主導。

That's the key differentiator from other AI chip startups.

這是與其他人工智能芯片初創公司的關鍵區別。

So your architecture uses high bandwidth memory?

Interesting, because this is, you know, high bandwidth memory is a key part of NVIDIA's products.

有趣的是，高帶寬內存是英偉達產品的關鍵部分。

As well from SK Hynix in particular.  But if you bring high bandwidth memory onto an inferencing chip, what kind of, what does that allow the chip to do?

尤其是 SK Hynix。  但是，如果把高帶寬內存引入推理芯片，芯片能做什麼？

And then ultimately, what does that mean for the AI application?

最終，這對人工智能應用意味著什麼？

Yeah, if we have higher bandwidth between the memory and our SOC, number one is we can get higher latency, token per second, for example, in NLM.

是的，如果我們在內存和 SOC 之間擁有更高的帶寬，第一，我們可以獲得更高的延遲，例如，在 NLM 中，每秒的令牌數。

Also, we have much less energy to get the data from the memory moving data, meaning the most important metric in here, token per second per watts, not just token per second.

此外，我們從內存移動數據中獲取數據所需的能量要少得多，這意味著這裡最重要的指標是每瓦每秒令牌數，而不僅僅是每秒令牌數。

Our architecture is highly optimized for high bandwidth memory, meaning that we need to have very scalable network on chip, on SOC, also memory subsystem, also modular design for the chip that we get higher latency, more throughput, most importantly, token per second per watts.

我們的架構針對高帶寬內存進行了高度優化，這意味著我們需要在芯片、SOC、內存子系統以及模塊化設計的芯片上建立極具可擴展性的網絡，從而獲得更高的延遲、更大的吞吐量，最重要的是每瓦每秒的令牌數。

So more power efficient as well is what you're saying.

所以你的意思是，它還更省電。

And that's obviously important when you think about the cost of running these models, right?

考慮到運行這些模型的成本，這一點顯然很重要，對嗎？

If you're a business and the cost is too high, you're not going to implement AI, right?

如果你是一家企業，成本太高，你就不會實施人工智能，對嗎？

So the cost needs to come down, power efficiency needs to be better.

是以，成本需要降低，能效需要提高。

And that's what you're saying that your products are working on.

這就是你所說的你的產品正在進行的工作。

字幕列表影片播放

與 Rebellions 公司首席執行官探討人工智能芯片的未來 (The future of AI chips with the CEO of Rebellions)

episode

multiple

basically

fascinating