Axelera Titania Chiplet Supports Native FP8 and FP4 In-Memory Compute
November 16, 2026
Axelera AI's Titania architecture uses digital in-memory compute and RISC-V cores to deliver 20 PFLOPS of FP8 in an 8-package, 4U air-cooled server. The system uses LPDDR6 instead of HBM for MoE decoding and operates at approximately 6 kW.
HOW THIS AFFECTS YOU
●
builderYou may see lower-cost inference options using LPDDR6 instead of expensive HBM for MoE models.
●
investorAxelera is positioning itself against major players like NVIDIA and AMD with specialized in-memory compute hardware.