Name: The future of AI chips with the CEO of Rebellions
Uploaded: 2025-07-25T18:36:23.000Z
Duration: 29 min 9 s
Description: Thousands of YouTube videos with English-Chinese subtitles! Now you can learn to understand native speakers, expand your vocabulary, and improve your pronunciation...

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

Participle clauses

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

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.

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

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

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

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

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

First of all, thank you for inviting me here.

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

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

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

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

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

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

What are your key products at the moment?

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

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

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

But let me talk about the functional diversities.

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

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

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

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

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

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

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

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

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

And NVIDIA has been very successful in this particular market.

Now, a lot of companies are talking about inferencing.

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

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

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

Let me back to your very first question here.

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

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

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

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

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

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

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.

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.

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

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

We don't need that much flexibility and luxury.

So it's exactly what deep sick was doing.

Deep sick developed their model using NVIDIA GPU.

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

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

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

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

Yes, without much fewer interconnects, much fewer SRAM.

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

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

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

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.

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

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

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

So your architecture uses high bandwidth memory?

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

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?

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

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.

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.

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.

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

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

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

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

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

Subtitles ListPlay Video

The future of AI chips with the CEO of Rebellions

episode

multiple

basically

fascinating