Rebellions Optimizes Chip For Latency

South Korean AI startup Rebellions has launched Atom, an AI-accelerator card targeting the high-frequency trading market. The company has raised over $100 million and has released ResNet-50 MLPerf benchmark performance data.

Anand Joshi

Korean startup Rebellions has introduced AI chips and accelerator cards targeting high-frequency trading (HFT) as a key application. The company is sampling its first product and expects to reach production in the second half of 2023. Rebellions has achieved a design win at Korea Telecom for cloud services unrelated to HFT.

Founded in 2020, Rebellions is headquartered in Seoul and has 80 employees. The company has raised more than $100 million in two funding rounds and boasts 28 PhDs on staff who graduated from top institutions such as MIT, Harvard, and KAIST.

The company’s first product is Atom, a PCIe accelerator card delivering 128 INT8 TOPS at 150 W TDP. The company uses Samsung’s 5 nm technology, and its primary focus is to minimize inference latency. The chip supports BF16, INT8, and INT4 data types. The company has released MLPerf Edge Inference results and beats Nvidia’s ResNet-50 latency score.

An AI chip company with a clear market focus and sound MLPerf results is a breath of fresh air in a crowded market. Rebellions can carve out a niche with HFT as a starting point. However, HFT is one of the many application markets for AI chips, and to become a formidable competitor, the company will have to improve its software and support multiple neural network architectures.