Axelera AI Unveils Titania AI Inference Architecture for Air-Cooled Data Center Racks

Related stories

Axelera AI unveiled details of the Axelera AI Titania architecture, its third-generation silicon platform built in a chiplet design for data center AI inference. According to Axelera AI, the architecture runs large models at interactive speeds within the power and cooling limits of existing data center facilities rather than requiring 100-kilowatt-plus liquid-cooled racks. Built for enterprises, sovereign clouds, modular data centers, telecommunications operators, and AI factories, Titania enters production in Q3 2027 with customer deployments scheduled through early 2028.

Why Power and Cooling Limits Shape Data Center AI Inference

TechTarget reports that AI inference requires different data center infrastructure than model training because inference workloads must process incoming user requests with low latency across varied rack densities and locations. According to TechTarget, citing Uptime Institute’s 2026 survey, more data center operators are reporting peak rack densities of 30 kW or higher, while high-density GPU racks such as Nvidia’s GB200 NVL72 draw approximately 120 kW at full load and require direct-to-chip liquid cooling.

TechTarget notes that air cooling remains suitable for CPU-based and lower-density accelerator configurations, enabling existing colocation facilities that cannot support dense GPU training clusters to host inference workloads. Bruce Bateman, chief analyst for semiconductors at Omdia, told TechTarget that inference offers an upgrade path for traditional data centers that cannot compete with large training facilities. The Axelera AI Titania platform pairs Digital In-Memory Computing (D-IMC) chiplets with LPDDR6 memory to operate within standard air-cooled server power envelopes.

How the Titania Chiplet Architecture and LPDDR6 Memory Work

Titania extends the D-IMC architecture introduced in Axelera AI’s Metis and Europa chips by adding floating-point support, more and smaller cores, and coordinated data movement. According to the Titania product page, each Titania system-in-package (SiP) houses two chiplets with 64 AI cores total, 32 RISC-V compute cores, and 128 MB of shared SRAM per chiplet, delivering 2.5 PFLOPS of FP8 compute at a 400 W thermal design power (TDP).

Replacing HBM with LPDDR6 Memory

Each Titania SiP incorporates 128 GB of LPDDR6 memory operating at 2,458 GB/s of memory bandwidth rather than high-bandwidth memory (HBM). Axelera AI states in its Titania launch announcement that sparse mixture-of-experts (MoE) models have large total parameter counts but small active parameter counts per token, making memory capacity and memory utilization more important than unused peak bandwidth. Using LPDDR6 provides 128 GB per package at lower memory power than HBM while avoiding HBM supply and packaging bottlenecks.

Also Read: Vertiv Launches Pre-Engineered SmartIT Platform to Streamline High-Density AI Data Center Deployments Worldwide

Simulated Token Throughput and Rack Power Density for Titania

Axelera AI expects Titania to achieve up to 7 times the throughput per watt of existing data center servers and up to 20 times better total cost of ownership (TCO). In October 2026 internal rack-level simulations at 200-plus tokens per second per user using 8k input tokens and 1k output tokens at FP4 precision, Axelera AI reports that an air-cooled rack with eight 4U Titania servers and 64 accelerators reaches 43,816 tokens per second on DeepSeek-R1 and 431,425 tokens per second on GPT-OSS-120B.

In those same October 2026 simulations, Axelera AI reports power efficiency of 1,053 tokens per second per kilowatt on DeepSeek-R1 and 10,371 tokens per second per kilowatt on GPT-OSS-120B. A single eight-SiP Titania server holds a 1.6-trillion-parameter model at FP4 precision with memory headroom for the KV cache, allowing two servers to fit inside a legacy 15 kW rack and five servers to fit inside a 30 kW air-cooled rack.

Titania AX8 Server Form Factors and Deployment Environments

Axelera AI plans to ship Titania in standard PCIe cards and Open Compute Project (OCP)-compliant UBB baseboards carrying eight Titania OAM modules linked in a scale-up network. The eight-SiP Titania reference server, codenamed AX8, delivers 20 PFLOPS of FP8 compute, 1 TB of accelerator memory, and 19.6 TB/s of memory bandwidth inside a standard 4U air-cooled chassis consuming 5 to 6 kW of typical power.

Enterprise Data Centers and Sovereign Clouds

Axelera AI’s product documentation lists enterprise data centers as a primary deployment environment, where organizations run large language models (LLMs), agents, and retrieval-augmented generation (RAG) on local data inside existing air-cooled racks. The product page also lists industry applications across language models, sovereign AI, agentic interactions, generative AI, computer vision, and medical and healthcare.

High-Performance Computing and AI Factories

Beyond enterprise server rooms, Axelera AI lists high-performance computing (HPC) centers, research institutions, and AI factories as target environments for Titania deployments. Operators in these facilities can deploy Titania accelerators to serve sparse MoE models while maintaining per-user token speeds as concurrent demand grows.

Voyager SDK and AxeleraScript Software Support Across Metis, Europa, and Titania

Titania runs on the same Voyager SDK and AxeleraScript (AxScript) software foundation used across Axelera AI’s first-generation Metis edge accelerator and second-generation Europa enterprise processor. Supported through the Voyager toolchain, Titania deployments include tensor, expert, and pipeline parallelism alongside speculative decoding, KV cache offloading, and disaggregation out of the box.

According to the Voyager SDK page, developers define end-to-end pipelines in YAML so the compiler can optimize and deploy neural networks alongside pre- and post-processing tasks across host CPUs, embedded GPUs, media accelerators, and Axelera AIPUs. The toolchain supports models from PyTorch, TensorFlow, and ONNX using TVM, GStreamer, and OpenCL/OpenGL compilers, and pairs with Axelera Wingman, an AI coding agent fine-tuned for Axelera hardware that builds, debugs, and optimizes inference pipelines from natural language prompts.

Axelera AI Commercial Commitments and European DARE Backing

Axelera AI reports more than $300 million in confirmed Titania commitments, with OEM partners preparing inference servers in both air-cooled and liquid-cooled configurations. The company has deployed its AI acceleration hardware across more than 600 customers in telecommunications, aerospace, and enterprise sectors, supported by its Partner Accelerator Network.

Titania was designed in Europe with backing from the EuroHPC Joint Undertaking’s DARE (Digital Autonomy with RISC-V in Europe) initiative, which operates under grant agreement No 101202459 to build European RISC-V-based HPC and AI hardware. Out of €240 million allocated by EuroHPC JU and the DARE consortium, Axelera AI received up to €61.6 million to develop the Titania chiplet architecture. “The biggest challenge most organizations face as inference becomes the dominant AI workload is the tokenomics of their infrastructure. For enterprises, it’s becoming impossible to own their own AI infrastructure due to high CAPEX expenditures including rebuilding their data center footprint,” said Fabrizio Del Maffeo, co-founder and CEO of Axelera AI.

Questions about Axelera AI Titania

Can existing Europa customers migrate to Titania without re-porting models?

Yes, the Titania PCIe card operates as a drop-in replacement for the Axelera Server 250p, a full-height full-length four-chip Europa card. Models deployed on Europa hardware today can move directly to Titania systems in 2028 without re-porting.

Which precision formats and video codecs does Titania support natively?

Titania natively supports MXFP4, FP4, MXFP8, FP8, and BF16 numeric data formats. Each card also includes on-chip 8K video decode and 4K video encode with H.265 and AV1 codecs to offload video pipelines from host CPUs and run language, speech, and vision models alongside live video.

Are Titania performance numbers based on silicon or simulations, and when does production start?

Performance figures published on Axelera AI’s product page are simulations run through Axelera AI’s production compiler rather than physical silicon measurements. Production is scheduled to begin in the second half of 2027, with customer deployments planned through early 2028.

What role do the 32 RISC-V cores play inside a Titania package?

Each Titania SiP integrates 32 programmable RISC-V compute cores across its two chiplets. These RISC-V cores handle collective communication operations across chips and offload business logic from the host CPU.

Technical Showcase at SC26 in Chicago

Axelera AI will present technical details of the Titania architecture at the SC26 supercomputing conference in Chicago, Illinois, from November 15 to 19, 2026, at booth 3251 in McCormick Place. Following the conference showcase, Titania moves toward Q3 2027 volume production and early 2028 customer deployments.

Subscribe

- Never miss a story with notifications


    Latest stories