Axelera Titania Brings Native FP8 and FP4 to In-Memory Compute
Source Entity
Hacker News

Axelera AI has unveiled its Titania chiplet, featuring native FP8/FP4 in-memory compute and RISC-V integration. The system promises high-density performance for AI workloads while utilizing LPDDR6 to reduce reliance on expensive HBM.
Axelera AI Unveils Titania: A New Frontier in In-Memory Computing
Axelera AI has officially introduced Titania, a groundbreaking chiplet architecture designed to redefine the efficiency of AI inference and training workloads. By integrating native FP8 and FP4 floating-point math directly into its digital in-memory compute (D-IMC) arrays, Axelera is tackling the primary bottleneck of modern AI hardware: the energy-intensive movement of data between memory and processing units. This architectural shift marks a significant evolution in how hardware handles matrix multiplication, the fundamental operation behind deep learning models.
The Power of D-IMC and RISC-V Integration
The third-generation D-IMC macro powering Titania represents a departure from traditional von Neumann architectures. By performing calculations within the memory array itself, Titania minimizes latency and power consumption. Furthermore, the inclusion of RISC-V orchestration and vector cores within every AI core provides a level of granular control that is essential for complex AI orchestration. This hybrid approach ensures that the hardware remains flexible enough to handle evolving model architectures while maintaining the raw throughput required for high-performance AI deployment.
Rethinking Memory Hierarchies: The LPDDR6 Shift
A critical aspect of the Titania announcement is the strategic choice to utilize LPDDR6 memory instead of the industry-standard High Bandwidth Memory (HBM). HBM has long been a cost and supply-chain bottleneck for AI accelerator manufacturers. By demonstrating the efficacy of LPDDR6 for Mixture of Experts (MoE) decode workloads, Axelera is positioning itself to offer a more scalable and cost-effective solution for data centers that are currently constrained by HBM shortages and the high thermal profiles of traditional GPU clusters.
Performance Density and Server Scalability
Axelera’s reference server specifications underscore the efficiency of this design. An 8-package, 4U air-cooled server configuration is projected to deliver 20 PFLOPS of FP8 performance while consuming only 6 kW of power. This density is particularly noteworthy when compared to heavyweights like the NVIDIA B200, AMD MI355X, or Qualcomm AI200. By prioritizing air-cooling over the complex liquid-cooling infrastructure required by many top-tier GPUs, Axelera is lowering the barrier to entry for high-density AI deployments.
Market Outlook and Production Roadmap
With $300 million in commitments already secured, the market interest in Titania is substantial. The company is targeting Q3 2027 for full-scale production, a timeline that aligns with the industry's anticipated demand for more efficient inference hardware. As the industry moves toward smaller precision formats like FP4 to accommodate larger models, Titania’s native support for these standards positions it as a highly relevant player in the future of AI infrastructure. Industry projections suggest that hardware efficiency will become the primary competitive differentiator as the cost of HBM continues to strain enterprise budgets.