Hacker News AIT2·Infra
74 pts
Axelera AI unveiled Titania, its third-generation inference chiplet that runs native FP8 and FP4 math inside digital in-memory compute arrays, with every matrix engine orchestrated by RISC-V cores and LPDDR6 replacing HBM for MoE decode. An 8-package, 4U air-cooled AX8 reference server delivers 20 PFLOPS of FP8 at roughly 6 kW, backed by $300 million in commitments ahead of Q3 2027 production.