[단독] 사피온과 합병 리벨리온 박성현 대표 "엔비디아보다 성능은 3배, 가격은 3분의1이 목표"

리벨리온 박성현 대표는 사피온과 합병 계획에 대해 치열한 글로벌 경쟁에서 대한민국을 대표할 AI반도체 기업으로 재탄생해 제대로 맞붙어보겠다고 했다.

박성현 리벨리온 대표가 지난 4월 분당 리벨리온 본사에서 진행된 코리아중앙데일리와의 인터뷰 후 포즈를 취하고 있다. 박상문 기자.

한국의 대표 영어신문, 코리아중앙데일리가 특종과 단독 인터뷰를 한글로 요약해 드립니다. 독자들의 많은 성원 바랍니다

최근 사피온과 합병 계획을 발표한 리벨리온의 박성현 대표에 따르면 글로벌 AI 반도체 시장은 이젠 기업 간의 경쟁을 넘어선 “초거대 자본이 개입된 싸움”이다.

시장을 독식하는 엔비디아를 뛰어넘기 위해 마이크로소프트, 오픈AI, 메타, 아마존, 애플 등 테크 강자들이 너나 할 것 없이 차세대 AI반도체 개발 경쟁에 뛰어든다. 리벨리온이나 사피온 같은 후발주자들이 합병을 통해 시너지를 모색하는 이유다.

박 대표는 코리아중앙데일리에 “이번 합병으로 두 회사가 대한민국을 대표할 AI반도체 기업으로 재탄생해 치열한 글로벌 경쟁에서 제대로 맞붙어볼 수 있도록 최선을 다하겠다”고 말했다.

향후 합병법인의 경영은 박 대표가 맡을 예정이다. 사피온의 류수정 대표는 사임한다. SK텔레콤 자회사인 사피온보다 스타트업으로 시작해 성공적으로 성장하고 있는 리벨리온 쪽에서 주도하는 편이 빠르게 변화하는 AI반도체 업계에 대응하는데 더 적합하다는 판단에 따른 것이다.

두 반도체 회사의 제품 포트폴리오 재편은 실사가 마무리된 뒤 결정될 전망이다. 하지만 리벨리온의 차기 반도체 제품인 리벨 (REBEL)은 그대로 유지될 것으로 보인다. 양산 직전 단계인 테이프 아웃 (tape-out: 최종 설계 도면이 대량 생산 라인으로 투입되는 단계) 시점이 9월로 예정돼 있기 때문이다. 합병법인의 기업공개(IPO)는 리벨리온 일정에 맞춰 추진할 것으로 알려졌다.

리벨리온은 올해 초 시리즈 B 투자 유치를 성공적으로 마무리하고 IPO 주관사를 선정하고 있다. 리벨리온의 누적 투자 유치 금액은 2800억원에 달하며 올해 초 기준 기업가치 8800억원으로 평가받았다.

다음은 지난 4월 리벨리온 분당 본사에서 박 대표와 진행한 인터뷰의 일문일답.

Q. AI 추론 영역에서, 경쟁사 제품 대비 리벨리온 반도체가 갖고 있는 강점은 무엇인가. A. 예를 들어, 메타의 LLaMA-65B 모델을 구동하려면, 엔비디아의 80 GB(기가바이트) 메모리가 탑재된 H100 GPU 두 장이 필요하다. 반면 9월에 테이프 아웃 예정인 우리의 리벨칩은 메모리 용량이 144GB로 거의 두 배이기 때문에 하나의 칩으로 이 모델을 구동할 수 있다.

가격은 어떻게 되나. 엔비디아와 비교해 성능은 3배, 가격은 3분의 1로 하는 것이 목표다.

리벨칩 개발과 생산에서 삼성전자와는 어떤 협업을 하나. 리벨칩에 삼성의 HBM3E를 배치하고 테일러 공장에서 4나노 공정을 활용해 제조하기로 했다. TSMC가 파운드리 1등인데 왜 삼성과 하는지 물을 수 있다. 물론 파운드리 공정은 TSMC가 잘하는 건 맞다. 하지만 LLM시장에서 파운드리만큼 중요한 것이 메모리다. TSMC엔 메모리 사업이 없다. 또 로직 웨이퍼와 메모리 웨이퍼를 연결해주는 패키징 기술도 필요한데, 삼성은 이 세 가지를 다 갖추고 있다. 우리처럼 작은 스타트업이 공급망을 개별적으로 다 따로 관리하는 것은 보통 힘든 일이 아니다. 삼성에 맡기면 턴키로 나오니까 종합적으로 봤을 때, 삼성전자와의 이런 협력 관계가 나쁘지 않다. 삼성 또한 우리를 통해 AI반도체 비즈니스 모델을 만들어내고 있어 이해관계가 맞는다.

패스트 팔로워 전략에 대해 더 설명한다면. LLM이라는 거대언어모델 등장 후 시장이 형성돼 있으니 여기에 제품을 맞춰 달리는 게임이 돼, 후발주자에겐 굉장히 편해졌다. 엔비디아가 GTC에서 로드맵을 말해주는 것은 우리에게 도움이 된다. 예전엔 어둠 속에서 달리는 느낌이었다. 프로덕트-마켓-핏(Product-Market Fit)이 틀어지면 엔비디아처럼 제품이 다양한 쪽이 이기는 게임이었지만, 지금은 마켓이 명확해져서 이에 맞는 제품을 만들 수 있는가, 없는가의 게임이다. 챗GPT로 정리됐고, 우리는 이 게임에 맞는 제품을 만들 수 있다. 불확실성이 사라진 상황이다.

리벨칩이 테이프 아웃을 앞두고 있는데, 시장 반응은 어떤가. 좋다고 할 수밖에 없을 것 같다. 우선 구글 같은 빅 테크 기업과 미팅을 하고 있고, 일본 통신사들과도 만나고 있다. 특히 IBM과 품질 테스트를 진행하고 있다는 점을 주목할만하다. IBM은 국가 안보와 관련된 일을 많이 하고 있어서 굉장히 보수적으로 테스트를 진행하기 때문에 이 자체로 매우 큰 의미다. IBM 품질 테스트를 통과하면 추후 다른 업체 품질 테스트를 좀 편히 갈 수 있어 초반에 IBM을 공략한 부분도 있다.

리벨칩의 초기 생산 분 수율은 어떠한가 수율을 정확히 알 순 없지만, 삼성전자가 “이 정도의 웨이퍼를 찍으면, 이 정도의 결과물이 나올 것이다”라고 예측한 것보다는 잘 나왔다.

━ 영어원문

In the Nvidia-dominated world of AI chips, a five-year-old startup from Korea, Rebellions, has some outrageous ideas to win the game.

Leveraging the benefits of a fast-follower, it is eying a shortcut to the top of AI inference, one of two major pillars in the AI ecosystem along with training. When that time comes, the idea of going shoulder-to-shoulder with Nvidia, in inference at least, won't sound so outrageous.

"In the past, before ChatGPT's emergence and before Nvidia's market dominance, it was like racing in the dark," says Park Sung-hyun, co-founder and CEO of Rebellions, an AI chip company based in Bundang, Gyeonggi, specializing in neural processing units (NPU) for AI inference, in an interview with the Korea JoongAng Daily.

Park, a native of the coastal city of Ulsan, did his master's and doctorate's degrees at the Massachusetts Institute of Technology before building his career as a chip expert at Intel, SpaceX and Morgan Stanley.

"But now the market has become clear with large language models (LLM) becoming the mainstream. And Nvidia is now presenting us with a road map. That guideline is a big help to small followers like us to quickly catch up in the market."

Park is putting up a strong fight against a coalition of big players like Nvidia, SK hynix and TSMC, while Rebellions relies solely on Samsung Electronics.

For small players like his company, however, he contends that having one partner that is able to handle everything is better as it saves resources.

"It's a turnkey with Samsung since it handles memory, manufacturing, and packaging," he says. "It is only one channel that we have to manage."

Park is confident that in AI inference, which prioritizes speed and cost efficiency over heavy computational skills in training, Rebellions can do better than others with its talented engineers and solid partners such as KT and Samsung Electronics.

His confidence is backed by hard evidence.

It recently closed on 165 billion won ($121 million) in Series B funding from global investors including Korea's KT as well as France's Korelya Capital and Japan's DG Daiwa Ventures.

The investment led its valuation to reach 880 billion won, leaving it only inches short of becoming a unicorn firm.

It was the only Korean company to make the list of the top 100 AI companies compiled by CB Insights that was released in April, joining OpenAI, Hugging Face, Anthropic and Groq among others.

This year is expected to be a major milestone for Rebellions because its ATOM NPU started mass production in March and a more premium REBEL NPU is slated for a tape-out in September with a plan for mass production next year in Pyeongtaek and Taylor, Texas.

A handful of other AI inference chips will flock to the market this year, two years on from the initial hype surrounding ChatGPT, and the real game to prove that Rebellions' silicon stands out among others will begin.

"We are going for the major league," Park says.

Below are the edited excerpts from the interview.

Q. In inference, can you give details on how Rebellions' chip is better than others? A. If you are to run Meta's LLaMA-65B model, you need two of Nvidia's H100 GPU, which has a memory capacity of 80GB. Our Rebel chip, which is set for a tape-out this September, has a memory capacity of 144GB. You only need one of our chips to run that language model compared to two of Nvidia's. What about the cost then? We have a target of offering an inference performance threefold that of Nvidia's at one-third of their price. Can you elaborate more on your fast-follower strategy? Before the advent of ChatGPT, AI was for only a few people. AI chip companies including us, in the middle of working on the ATOM chip for finance and small data centers, ended up having the product-market-fit all messed up because we didn't know that AI services like ChatGPT, which are for the masses and require a huge amount of data processing, will become mainstream. Nvidia also did not expect it, but due to its expansive portfolio, it was able to respond to the market well. Before, it was a game that companies with big portfolios were destined to win. Now, with the market narrowed down, it is a game of who can make a product that perfectly fits the market. And we are saying we can make it, and we can make it better than others. Did you have to start the design and business strategy all over again with the advent of ChatGPT? Actually it was an extension of what we were doing. What we prioritized was latency when we were focusing on chips for the finance market. In inference for LLMs, latency, an indicator of how fast the model coughs up its first response, which we call the first token, is crucial as well. So it was a natural flow. You closed Series B funding early this year with 165 billion won, which is a notable feat considering not-so-favorable market conditions. Now your accumulated investment adds up to 277 billion won. What was the process like? We were able to persuade investors that Rebellions is the top player for AI chips in Korea with our talent pool. One of our founding members is Dr. Oh Jin-wook, our CTO who is from IBM and has a global reputation. In fact, although I am sitting here doing the interview, when abroad, foreign investors call us Jin-wook's team. Another founding member is CPO Dr. Kim Hyo-eun, who was already respected in this technology scene, having served as CPO at Lunit. Kim's reputation goes as far as being able to start a company of his own now and still being able to raise funds. In the AI chip scene, it is impossible to make an investment decision after the silicon is out in the market because it requires a tremendous amount of money and also time. Investors make decisions first from a human resource perspective, and then in a second round with the product. Our strategic partner KT also played a big part. KT, Korea's second-biggest telecom company, invested a total of 66.5 billion won into Rebellions so far and is its strategic partner, owning at least 10 percent of the startup. Rebellions has also provided its ATOM chips to KT's data center to run its cloud services. KT doesn't perceive AI as some cool, high-tech algorithm. It sees it as an infrastructure. They plan on building a range of data center infrastructure to enable AI. They are good at building infrastructure, having done so with 3G, 4G and 5G networks. KT is actually the biggest operator of internet data centers in Korea. When the investors carried out CDD, or commercial due diligence, the fact that a local company has a local data center was perceived well. KT's strategic investor position also served as a kind of shield because the investors would think that with such a big company behind us, Rebellions will never go out of business. The Rebel chip, obviously your main product since it is named after the company, is scheduled for a tape-out in a few months. What has the initial response from the market been like? It is positive. We are having meetings with Google and other Big Tech and telecom companies from Japan. We are also running a qualification test with IBM, which is known for taking a very conservative approach to such tests. This is going to take a while because they are running aging tests as well to determine the chips' durability. One of the criteria they look at closely is the product's reliability. They have a very high standard because their systems are used by the military as well as in enterprise and public finance. If we qualify at a certain level in IBM's test, it would allow us to take an easier path with other tech companies, which is why we have targeted IBM first. You have a close partnership with Samsung Electronics. What has that been like? Yes, we are deploying their HBM3E (the next-generation high bandwidth memory) chips on our Rebel and we are contracting its manufacturing at Samsung's Taylor, Texas factory utilizing its four-nanometer process. People may cast doubt because with foundries (contract manufacturing), TSMC is as you know No. 1. But as much as foundry is important for AI chips, memory chips are also crucial. TSMC doesn't have it. And it doesn't end there. Chiplet packaging technology that connects all the different components together is also important. And you know what? Samsung Electronics has it all. For a small startup like us, it is difficult to manage a supply chain involving different companies like HBM from SK hynix, packaging from Amcor and foundry with TSMC. In that sense, Samsung Electronics is a good partner. Did the Rebel chip's early production phase with Samsung go well, such as the yield? It is hard for us to know the yield, but the outcome matched what they told us, such as if you put in a certain amount of wafer you will get certain amount of output. Actually, the outcome was better than expected because their PDK (process design kit; a set of files delivered by chip manufacturers to the designers about the fabrication process) has been set very conservatively. They don't exaggerate it. Perhaps that's because they don't have that many clients yet. If they exaggerate the PDK, they will not be able to secure clients in a long term. What is your plan with going public? We are going there for sure and we have started preparations. We are going to list on the Korean market first and head to the U.S. for dual listing. We plan on running in the "K-league" first until our valuation reaches somewhere near $4 billion and then move to the major league in the U.S.

BY JIN EUN-SOO, PARK EUN-JEE [jin.eunsoo@joongang.co.kr]