Axelera AI Unveils Titania Data Center AIPU Architecture, Targeting 2027 Production
Axelera AI announced its third-generation Titania AIPU architecture aimed at the data center inference market, with mass production slated for the third quarter of 2027. The company has already secured over $300 million in orders for the new chip. Titania promises high energy efficiency and lower total cost of ownership, offering air-cooled 4U servers that can compete with top-tier liquid-cooled systems like NVIDIA's Vera Rubin NVL72. This development highlights the growing competition in specialized AI inference hardware and memory integration. The 400W Titania chip features a dual-chiplet design based on the third-generation D-IMC architecture, delivering 2.5 PFLOPS of FP8 compute and integrating 128GB of LPDDR6 memory via SiP. It also includes 32 RISC-V cores to offload host CPU tasks and supports both PCIe AIC and OCP OAM form factors.
## BACKGROUND
AIPU refers to an AI Processing Unit, a specialized microprocessor designed specifically to accelerate artificial intelligence workloads. Data center inference is the phase where a trained AI model processes live input data to make predictions or decisions, which typically requires high throughput and energy efficiency. Technologies like LPDDR6 memory and OCP OAM form factors help achieve high-density, high-bandwidth computing performance in modern server infrastructures.