一家芯片新贵，组团对抗英伟达

Core Insights - The article highlights the emergence of Rebellions AI, a South Korean startup that has gained significant backing from major HBM memory manufacturers and telecom companies, positioning itself to create AI accelerators for data centers [1][9] - Rebellions aims to leverage its partnerships and advanced chip technology to compete in the AI inference market, particularly focusing on the flexibility and adaptability of its second-generation AI accelerators [2][3] Company Overview - Rebellions was founded in September 2020, initially targeting AI inference chips for high-frequency trading, and has since evolved to compete with established players like NVIDIA and AMD [3][8] - The company is headquartered in Seoul, South Korea, which is a significant industrial and financial hub, projected to reach a GDP of $1.86 trillion by 2025 [5] Founders and Leadership - The company was co-founded by four individuals, including CEO Park Sung-hyun, who has a strong background in engineering and finance, and CTO Woo Jin-seok, who has extensive experience in AI research [5][7] Funding and Valuation - Rebellions has successfully raised $61 million in two rounds of funding and became South Korea's first AI chip unicorn with a valuation exceeding $1 billion after merging with Sapeon Korea [8][9] Technology and Product Development - The company has transitioned from using TSMC's 7nm process for its Ion chip to Samsung's 4nm process for its Rebel series chips, which are designed to compete with data center-level GPU accelerators [10][11] - The Rebel AI inference chips utilize a Coarse Grained Reconfigurable Array (CGRA) architecture, allowing for programmable interconnections between processing units, enhancing flexibility for various AI tasks [11][13] Performance Metrics - The Rebel Single chip is designed to achieve 160 teraflops at FP16 precision and 320 teraflops at FP8 precision, with a clock frequency estimated around 2 GHz [16][30] - The Rebel Quad configuration can deliver 1 petaflops at FP16 and 2 petaflops at FP8, with a power consumption of 600 watts, making it competitive with NVIDIA and AMD offerings [27][30] Market Positioning - Rebellions is strategically entering the AI inference market, which is seen as a lucrative area for monetization, especially as the demand for tensor mathematics and HBM memory exceeds supply [31][34] - The company is developing a software stack based on open-source technologies to support its hardware, including a communication library similar to NVIDIA's NCCL [34]