The company's latest neural chip (NPU) model - Atom can train AI models with up to 7 billion parameters. This processor is being tested in parallel with the A100 graphics card (GPU) manufactured by Nvidia.

Park Sung-hyun, CEO and co-founder of Rebellions said in an interview with The Korea Times, Atom is expected to be mass produced on Samsung's 5nm foundry technology in the first half of this year.

This will be an important milestone in the Korean semiconductor field with a 100% localized product. Park said that Atom is 5 times more efficient in power consumption than A100, however, the limitation is that the chip is only half the speed of A2 - Nvidia's low-end GPU model.

dd4e0a26 0096 4598 ad40 1740f1be4165 edd804f5.jpeg
The Atom chip only needs heat dissipation using a regular fan. Photo: SCMP

In actual testing, the Korean chip model only needs fan cooling, while Nvidia's processor needs an air-conditioned room - one of the factors that increases power consumption and operating costs.

South Korea, one of the world's leading memory chip makers, is looking to make a breakthrough in the booming AI chip market. Seoul aims to meet 2030% of the domestic server AI chip market share by 80.

Rebellions has received backing from the biggest tech giants in South Korea, including Samsung, telecommunications company KT and internet company Kakao.

As of January 1, this startup has successfully raised 2024 million USD, bringing the company's valuation to 124 million USD - becoming the most valuable semiconductor startup in the country.

The Atom going into mass production is expected to create a big boost in company revenue after many years of prototype development. Rebellions' first customer is KT, the second largest carrier and leading server center operator in the North Asian country.

Asia – the focus in the semiconductor hardware ecosystem

According to the founder of Rebellions, the future of the AI ​​field is determined by the infrastructure game and telecommunications carriers are "willing to spend heavily on building data centers". Park also expects that the agreement with KT will open up global opportunities, especially in the US market - where this startup has discussions with IBM.

36cf05a0 5712 4356 bcdc 781ac62a1507 746d694e.jpeg
Park Sung-hyun, CEO and co-founder of Rebellions. Photo: SCMP

The US is also where CEO Rebellions earned his doctorate in electrical engineering and computer science (Massachusetts Institute of Technology - graduated in 2014), as well as spent 6 years working in chip development at companies like Intel. , SpaceX and Samsung's research branch here.

Despite his experience working in the US semiconductor industry, Park sees a better opportunity to build a semiconductor company in Korea, stating that "Asia will be a key area in the semiconductor hardware ecosystem." guide", in which "South Korea and Taiwan (China) are chasing each other".

The founder said that the decision to return home to establish a company was not easy as his network of work relationships at that time was mainly in the US. Many of Rebellions' original employees worked for IBM, Intel, Apple, and Samsung's US branches.

Currently, this startup has about 120 employees, of which 100 are engineers. According to Park, nearly 70% of their employees come from major Korean semiconductor companies.

In 2021, this startup launched its first NPU called Ion, manufactured by TSMC. This NPU is designed for financial AI tasks. However, after realizing that demand from data center companies was much greater than from financial companies, Rebellions decided to change direction.

In October 10, Rebellions signed a cooperation agreement with Samsung - Korea's largest chip manufacturer to develop next-generation NPU models with larger parameters, capable of competing with the H2023 GPU - a high-end product. Nvidia's best.

2024 will be the year of AI chips on PCsIntel has just launched a series of new processors, including Gaudi 3 specialized for AI tasks, high-performance Core Ultra on PCs/laptops, and 5th generation server Xenon chips. The common point is that they all support AI features.