Axelera AI’s Titania is a third-generation inference chiplet designed to bring digital in-memory compute into data centers. Unlike its Europa predecessor, which converted floating-point models to integer formats, Titania performs MXFP4, FP4, MXFP8, FP8, and BF16 calculations directly in its SRAM arrays. Each chiplet combines 32 AI cores, 16 compute cores, and 128 MB of shared SRAM, delivering a claimed 1.25 PFLOPS of FP8 performance. Two chiplets form a 400-watt package with LPDDR6 memory, while the AX8 reference server packs eight packages into an air-cooled 4U chassis. RISC-V orchestration and vector processors inside the compute blocks are intended to make each AI core a programmable inference tile.
The design targets large-model inference, particularly mixture-of-experts workloads, with LPDDR6 offering an alternative to costly HBM. Axelera says an AX8 can hold a 1.6-trillion-parameter model in FP4, and claims up to seven times the throughput per watt and 20 times better total cost of ownership than leading servers. It reports more than $300 million in commitments and expects production to start in Q3 2027, with volume arriving in 2028. The key question is whether Titania’s new floating-point macro will deliver its simulated performance in silicon: the chip has not taped out, and the comparisons use GPU figures from September 2026.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.