📅 November 12, 2026 • Source: Axelera AI
Axelera AI has unveiled Titania, its third-generation silicon architecture after Metis for the edge and Europa for the enterprise. Per the company, production is set for Q3 2027 with deployment commitments through early 2028. Titania is a chiplet design built on Digital In-Memory Computing with programmable RISC-V compute, LPDDR6 memory and high-bandwidth scale-up connectivity, targeting data-center-grade inference in air-cooled racks instead of 100-kilowatt-plus liquid-cooled GPU systems.
Each Titania system-in-package carries two chiplets and delivers 2.5 PFLOPS of FP8 at 400 W, per Axelera AI. Per Futurum, that is 1.25 PFLOPS FP8 per chiplet with 32 AI cores each, built on TSMC N4C with a UCIe interconnect and a PCIe Gen6 host link. Memory is LPDDR6 instead of HBM: 128 GB per package at 2,458 GB/s per the product page, configurable up to 2 TB per package. The silicon also carries on-chip media engines with 8K decode and 4K encode in H.265 and AV1, per Axelera AI.
The AX8 reference server packs eight SiPs into a standard 4U air-cooled chassis, per Axelera AI: 20 PFLOPS FP8, 1 TB of accelerator memory and 19.6 TB/s of bandwidth at a typical 5-6 kW. The company says two servers fit a legacy 15 kW rack and five fit a 30 kW air-cooled rack, with one server holding a 1.6-trillion-parameter model at FP4 with KV-cache headroom, per Axelera AI. The company frames this as HPC-class responsiveness without liquid-cooled racks.
Software remains the same Voyager SDK and AxeleraScript across generations, with tensor, expert and pipeline parallelism, speculative decoding and KV cache offloading. Per Axelera AI, the design promises up to 7x throughput per watt, up to 20x better TCO and over $300 million in confirmed commitments. CTO Evangelos Eleftheriou says it already turns 88 percent of memory bandwidth into tokens on Europa silicon that is shipping today. Futurum balances: the proof sits in simulations before tape-out, with volume in 2028 — claims stay projections until silicon ships.
IMPACT: Titania aims inference capacity at the air-cooled racks enterprises, sovereign clouds, telcos and neoclouds already run — no liquid-cooled rebuild required. LPDDR6 instead of HBM sidesteps HBM cost, and per Axelera AI the shared SDK path, including a drop-in PCIe card for the Axelera Server 250p, lets models move to Titania in 2028 without re-porting. A technical session at SC26, November 15-19 in Chicago, booth 3251, covers the architecture details. For buyers, silicon validation is the milestone to watch next.
Sources:
https://axelera.ai/news/axelera-ai-unveils-titania-architecture-hpc-class-ai-responsiveness-without-liquid-cooled-racks-targeting-110b-underserved-market
https://techpowerup.com/354736/axelera-ai-unveils-titania-architecture
https://www.hpcwire.com/off-the-wire/axelera-ai-unveils-titania-architecture-ahead-of-sc26/
https://futurumgroup.com/insights/axelera-titania-brings-native-fp8-and-fp4-to-in-memory-compute
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
