Conference Presentation, Product Demonstration, Keynote
June Paik (FuriosaAI): Tensor Contraction Processor: NextGen AI Inference Chip for Data Centers
- Furiosa AI aims to develop native silicon chips for AI data center inference, anticipating a paradigm shift from training to inference as AI becomes pervasive across industries and personal life within the next five years.
- The company projects unprecedented compute demand driven by agentic AI applications capable of reasoning and planning, necessitating thousands to hundreds of gigawatt-scale data centers fueled by massive capex and sovereign AI initiatives.
- Supplying the estimated 100 gigawatts of energy required for future data centers is identified as a major challenge, as this volume matches the peak consumption of an industrialized nation like Korea.
- The company prioritizes economic and energy sustainability, noting that business models will fail if AI operational costs are too high, and seeks to replicate the efficiency transition from gasoline to electric vehicles in the energy sector.
- The Renegade chip is designed to consume 180 watts, compared to over 300 watts for equivalent GPUs, with projections indicating that saving 100 watts per chip yields $20 in annual savings and could save nearly $10 billion over five years for a 100-megawatt data center.
- Manufacturing utilizes a 5-nanometer node from TSMC and advanced memory integration with SK Hynix to pack 40 billion transistors into a single system-on-chip die.
- Benchmarking indicates the Renegade chip outperforms the H100 by nearly 190% in real-world LLM workloads when measured in tokens per second per watt.
- Hardware design requires a holistic co-design of algorithms, hardware, and software to avoid efficiency losses associated with rigid specific optimizations that cannot adapt to rapid innovation.
- Raising the hardware abstraction level to tensor contraction is expected to make complex NP-hard computing metrics more tractable and streamline software development.
- The company is currently sampling with global enterprises and anticipates deploying applications across manufacturing, life sciences, energy, and the public sector, including sovereign AI contexts.
- Furiosa AI urges enterprises to own their AI stacks and computing layers to maintain cost control, warning that reliance on external solutions prevents cost management.
- The Renegade represents the second-generation product from a Seoul-based startup that has been developing technology for eight years, with the stated ambition to lead the AI computing space.