The episode of the AI Hardware Podcast focuses on the burgeoning AI chip industry in South Korea, highlighting several key companies and their unique approaches to AI hardware. The discussion begins with DeepX, a relatively young but prolific company producing multiple chip series tailored for computer vision applications, particularly in industrial and security camera markets. Despite modest revenue figures reported in 2023, DeepX benefits from strong partnerships with Korean telecom firms and access to advanced manufacturing processes like Samsung’s 5-nanometer technology. Their collaboration with Baidu to integrate with the PaddlePaddle AI framework signals ambitions to expand into the Chinese market.
Next, the podcast covers Hyper Excel, a startup aiming to compete with established players like Groq by focusing on low-latency, high-throughput AI inference for smaller language models. Their upcoming large ASIC, named Bertha, built on Samsung’s 4-nanometer process, targets efficient deployment of models with one to two billion parameters, suitable for fast, imprecise decision-making tasks. Although still early in funding stages, Hyper Excel’s approach reflects a trend toward specialized AI chips optimized for specific workloads, contrasting with general-purpose GPUs.
Mobileint, another Korean startup, specializes in AI chips designed for edge devices such as drones and smart retail environments. Their chips, Ares and Regulus, focus on convolutional neural network workloads for computer vision, emphasizing local processing to reduce cloud dependency. Mobileint also supports major AI frameworks like PyTorch and TensorFlow, enhancing software compatibility. While still building their customer base, their hardware includes features like video codec acceleration and low-light performance, catering to industrial automation and robotics sectors.
The discussion then shifts to two major Korean players: Furiosa and Rebellions. Furiosa, known for its Warboy and second-generation Renegade chips, targets enterprise AI markets with high-performance inference accelerators featuring HBM3E memory. Despite turning down a significant acquisition offer reportedly from Meta, Furiosa continues to develop independently, securing clients like LG AI Research. Their technology focuses on tensor compression to optimize memory and compute efficiency, positioning them as a strong contender in data center AI hardware.
Finally, Rebellions, formed from a merger with Zappion, develops chiplet-based AI processors designed for telecommunications cloud applications. Their Rebel chip features multiple tensor processors, high memory bandwidth, and integration with the ARM chiplet ecosystem, enabling flexible deployment scenarios. With a clear roadmap extending to 2027 and ongoing collaborations, Rebellions aims to capture a significant share of the AI inference market. Overall, the episode underscores South Korea’s growing AI chip ecosystem, driven by diverse startups and established firms leveraging local partnerships and advanced semiconductor technologies to compete globally.
Useful Links
