Axelera AI details Titania inference chip with native FP8/FP4, targets Q3 2027 launch
Axelera AI disclosed the architecture of Titania, its third-generation inference chiplet, which performs FP8 and FP4 math directly within digital in-memory compute arrays and pairs every matrix engine with RISC-V cores. An 8-package AX8 server built on two-chiplet SiPs is said to deliver 20 PFLOPS of FP8 compute and up to 1 TB of LPDDR6 memory while drawing about 6 kW, with production slated for Q3 2027 backed by over $300 million in confirmed commitments.
GoKawiil's interpretation of the reporting above, not reported fact.
By using LPDDR6 instead of HBM and keeping the system air-cooled, Axelera appears to be positioning Titania as a lower-cost, lower-power alternative to GPU-based inference servers from Nvidia, AMD, Intel and Qualcomm. The company's own projections of up to 7x better performance-per-watt and 20x improved TCO are self-reported figures, so independent benchmarking would be needed to confirm whether Titania can match rival systems once deployed at scale in 2027.
- Titania chiplets combine native FP8/FP4 in-memory compute with RISC-V cores on TSMC N4C.
- An 8-package AX8 server targets 20 PFLOPS FP8 and 1TB LPDDR6 at roughly 6kW.
- Axelera has secured over $300 million in commitments ahead of Q3 2027 production.
Source: futurumgroup.com — Brendan Burke, 2026-11-16
Published there as: “Axelera Titania Brings Native FP8 and FP4 to In-Memory Compute”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.