Axelera Titania Brings Native FP8 and FP4 to In-Memory Compute (futurumgroup.com)

🤖 AI Summary
Axelera AI has announced Titania, its innovative third-generation chiplet designed for digital in-memory compute (D-IMC) that introduces native FP8 and FP4 mathematical capabilities. With each Titania chiplet featuring 32 AI cores, RISC-V orchestration, and 128 MB of shared SRAM, the architecture promises significant performance improvements, delivering 1.25 PFLOPS of peak FP8 compute. In a system package utilizing eight chiplets, the compute performance scales to an impressive 20 PFLOPS while maintaining energy efficiency at around 6 kW. Axelera's projections indicate Titania could provide up to 7x throughput per watt and 20x better total cost of ownership compared to existing data center servers, marking a substantial technological leap in AI inference capabilities. The significance of Titania within the AI/ML landscape lies in its capability to handle floating-point data natively, potentially streamlining the deployment of complex models like GPT and Qwen without the need for integer requantization required by previous generations. Additionally, the integration of RISC-V cores optimizes data handling across the AI cores, allowing each to function as a programmable unit. Titania's design also favors air-cooled server infrastructures, catering to existing data center setups, which significantly positions it as a viable solution for enterprises looking to modernize their AI compute capabilities without the need for extensive infrastructure overhauls. Production is set to begin in Q3 2027, with over $300 million in commitments already secured.
Loading comments...
loading comments...