Conference Presentation, Product Demonstration
Rethinking AI Infrastructure for the Age of Reasoning | Jeehoon Kang, FuriosaAI | RAISE 2026
- The industry is expected to expand data center construction with a majority of new facilities dedicated to AI inference workloads, while power availability remains the primary constraint for the sector.
- Furiosa intends to address power limitations by developing power-efficient AI chips and systems capable of generating 30% to 100% more tokens per rack compared to competitors.
- Second-generation Renegade chips are scheduled for mass production this year with an expected output of 20,000 units, featuring a hybrid HBM-SRAM architecture and a 180-watt TDP.
- Renegade-based servers are designed to consume only 3 kilowatts, allowing five to six units to fit in standard air-cooled racks, with orders currently accepted and delivery expected within days.
- The Renegade utilizes a tensor contraction processor (TCP) architecture that eliminates dynamic features to ensure performance predictability, supported by a compiler that simplifies memory and instruction selection through "shapes" and "tactics."
- Development is underway on a third-generation "Torque" chip in partnership with Broadcom, designed as a scalable hyperscale accelerator for direct liquid cooling environments.
- The Torque chip aims to deliver 10 to 30 times the capacity of the Renegade chip and is scheduled to demonstrate integration with VoxTraw and Mistral models.
- The company positions its products as ready for immediate deployment for customers requiring more power-efficient AI processing than GPUs.