news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Oh Jin-wook, chief technology officer of South Korean AI chip design company Rebellions, said in an interview that its next-generation AI chip REBEL is expected to be released in 2024.

REBEL is designed for training large language and multimodal models, and uses Samsung wafer manufacturing technology and memory chips. In terms of process technology, it is expected to use Samsung's 4-nanometer node process, while the memory chip is mainly Samsung's HBM3E memory chip. In addition, the REBEL chip will also support 800Gb Ethernet transmission interface.

At present, the REBEL family includes two products, namely the single-chip REBEL-Single and the four-chip REBEL-Quad, which are equipped with one and four 12-layer stacked HBM3E memory chips respectively. For REBEL-Quad, it can train large language models with a parameter size of 175B+.

Oh Jin-wook said the company plans to release REBEL-Single by the end of 2024.As for the larger REBEL-Quad, thanks to the strong support of Samsung Electronics, its launch time is expected to be advanced from around 2026 to early 2025.

In fact, Rebellions signed a final merger agreement with SAPEON Korea, another AI semiconductor design company under the Korean SK Group, on the 18th of this month. After the two companies completed the merger, SAPEON Korea became the surviving company, but its name was changed to Rebellions. The new Rebellions will be led by the current CEO Park Sung-hyun, and the enterprise value will exceed 1 trillion won (about 740 million US dollars).

According to previous reports, the new Rebellions, led by CEO Park Sung-hyun, aims to seize the global AI semiconductor market within two to three years, challenge NVIDIA's position, and focus on the research and development of NPU. However, the market previously estimated that the new Rebellions might adopt TSMC's process technology. However, now the new Rebellions has come out to refute the rumor, saying that the new AI chip will be manufactured using a 4-nanometer node process.