Titania
Titania is a high-performance, energy-efficient AI inference chiplet developed by Axelera AI, designed to deliver server-grade performance with the power efficiency required at the edge.
Publisher review
Titania is a high-performance, energy-efficient AI inference chiplet developed by Axelera AI, designed to deliver server-grade performance with the power efficiency required at the edge. It is engineered for high-performance computing (HPC) environments, targeting researchers and enterprises needing scalable AI inference for workloads like climate modeling, drug discovery, and predictive analytics. Backed by up to €61.6 million from the EuroHPC Joint Undertaking and member states under the DARE Project, Titania aims to provide a modular, scalable AI inference platform starting in 2027, addressing Europe's need for sovereign AI hardware supply chains.
Titania is built on Axelera AI's proprietary Digital In-Memory Computing (D-IMC) architecture, which integrates compute and memory to reduce data movement and power consumption. It features proprietary RISC-V vector extensions for diverse AI workloads and supports near-linear scalability from edge to cloud via a chiplet-based design. The chiplet can be combined into PCIe boards and OCP-like insert cards, enabling configurations from single-chip edge devices to multi-chip HPC clusters. This architecture delivers high throughput and low latency for inference tasks, with performance scaling linearly as chiplets are added.
Titania competes directly with Nvidia's GPU-based inference solutions, which dominate the HPC market but consume significantly more power per operation. It also competes with Openchip and Codasip, which offer RISC-V-based AI accelerators but lack Titania's D-IMC approach and chiplet scalability. Axelera's focus on energy efficiency and modularity positions Titania as a viable alternative for European HPC centers seeking to reduce reliance on non-European suppliers, though it faces an uphill battle against Nvidia's established software ecosystem and market inertia.
The honest trade-offs include high development costs due to the complex chiplet architecture, which requires advanced packaging and interconnects. The path from small-scale prototypes to large-scale production is long and risky, with Titania not expected until 2027. While the D-IMC architecture offers power efficiency, it may not match Nvidia's raw floating-point performance for training workloads, limiting its use to inference only. Additionally, the RISC-V ecosystem is less mature than Nvidia's CUDA, potentially requiring custom software optimization for some models.
How it works
-
Digital In-Memory Computing
Proprietary D-IMC architecture integrates compute and memory to minimize data movement, reducing power consumption while maintaining high inference throughput.
-
Chiplet-based scalability
Modular chiplet design enables near-linear performance scaling from single edge devices to large HPC clusters, supporting PCIe and OCP form factors.
-
RISC-V vector extensions
Enhanced with proprietary RISC-V vector instructions to accelerate diverse AI workloads, including convolutional and transformer models.
-
Purpose-built for inference
Optimized specifically for AI inference tasks, delivering high throughput and low latency for applications like real-time video analysis and scientific modeling.
-
Edge-to-cloud deployment
Designed for deployment across edge, enterprise data center, and HPC environments, with consistent performance scaling across form factors.
-
EuroHPC funding support
Backed by up to €61.6 million from the EuroHPC Joint Undertaking and member states, ensuring alignment with European digital sovereignty goals.
-
Modular platform design
Supports PCIe boards and OCP-like insert cards, allowing system integrators to configure compute density based on workload requirements.
Strengths and trade-offs
Strengths
- Delivers server-grade AI inference performance while consuming only a fraction of the power of comparable Nvidia GPUs, thanks to its D-IMC architecture.
- Near-linear scalability from edge to cloud allows a single chiplet design to be used in devices ranging from compact edge nodes to large HPC clusters.
- Backed by up to €61.6 million in EuroHPC funding, providing substantial financial support for development and reducing time-to-market risk.
- Proprietary RISC-V vector extensions enable efficient processing of diverse AI models, including CNNs and transformers, without relying on proprietary instruction sets.
Trade-offs
- High development costs due to the complex chiplet architecture, which requires advanced packaging and interposer technologies.
- Long and risky path from small-scale prototypes to large-scale production, with availability not expected until 2027.
- Limited to inference workloads, unable to compete with Nvidia's GPUs for AI training tasks that require high floating-point performance.
- RISC-V software ecosystem is less mature than CUDA, potentially requiring significant custom optimization for some AI models and frameworks.
Pricing context
Not specified; expected to be a high-cost, enterprise-grade chiplet solution for HPC and data center deployments, with pricing likely per chiplet or per board.
Getting started with Titania
-
Request early access
Visit the Axelera AI website and submit an inquiry for early access to the Titania chiplet development program. Provide your organization details and intended use case, such as HPC inference for climate modeling or drug discovery.
-
Set up development environment
Install the Axelera AI SDK and toolchain on your Linux workstation. Follow the provided documentation to configure the RISC-V cross-compiler and link the D-IMC libraries for your target AI models.
-
Load and optimize models
Convert your trained AI models (e.g., PyTorch or TensorFlow) to the supported format using the Axelera model optimizer. Apply quantization and pruning to leverage the D-IMC architecture's power efficiency.
-
Run inference on chiplet
Deploy the optimized model to a single Titania chiplet on a PCIe board. Execute a test inference workload, such as real-time video analysis, and measure throughput and latency using the provided monitoring tools.
-
Scale to multi-chiplet cluster
Connect multiple Titania chiplets via the chiplet interconnect to form a larger inference cluster. Configure the system software to distribute workloads across chiplets, verifying near-linear performance scaling.
Frequently Asked Questions
What is the Titania AI chiplet and who makes it?
Titania is a high-performance, energy-efficient AI inference chiplet developed by Axelera AI. It uses proprietary Digital In-Memory Computing to reduce power consumption and is designed for scalable deployment from edge devices to HPC clusters.
How does Titania's Digital In-Memory Computing improve efficiency?
Titania's D-IMC architecture integrates compute and memory to minimize data movement, which significantly reduces power consumption while maintaining high inference throughput. This allows server-grade performance at a fraction of the power used by comparable Nvidia GPUs.
Can Titania scale from edge devices to large HPC clusters?
Yes, Titania uses a modular chiplet design that supports near-linear performance scaling. It can be combined into PCIe boards or OCP-like cards, enabling configurations from single-chip edge devices to multi-chip HPC clusters with consistent performance.
How does Titania compare to Nvidia GPUs for AI inference?
Titania delivers competitive inference performance with much lower power consumption than Nvidia GPUs. However, it is limited to inference tasks and cannot match Nvidia's floating-point performance for training. It also relies on a less mature RISC-V ecosystem versus CUDA.
What funding supports the development of Titania?
Titania is backed by up to €61.6 million from the EuroHPC Joint Undertaking and member states under the DARE Project. This funding supports development and aligns with Europe's goal of creating a sovereign AI hardware supply chain.
When will Titania be available and what are its main limitations?
Titania is expected to be available starting in 2027. Its main limitations include high development costs due to complex chiplet architecture, a long path to production, and restriction to inference workloads only, with a less mature software ecosystem than CUDA.
Alternatives
How Titania compares
Direct head-to-head against 3 competitors. Picked by 7wData.
Titania
- Pricing
- Not specified; expected to be a high-cost, enterprise-grade chiplet solution for HPC and data center deployments, with pricing likely per chiplet or per board.
- Target
- Titania is a high-performance, energy-efficient AI inference chiplet developed by Axelera AI, designed to deliver server-grade performance with the power efficiency required at the edge.
- Strength
- Delivers server-grade AI inference performance while consuming only a fraction of the power of comparable Nvidia GPUs, thanks to its D-IMC architecture.
- Watch for
- High development costs due to the complex chiplet architecture, which requires advanced packaging and interposer technologies.
Tenable Nessus
- Pricing
- Starts at $2,990/year per scanner; Professional $3,990/year
- Target
- Vulnerability assessment teams needing broad, accurate scanning across on-prem and hybrid environments
- Deployment
- On-prem, cloud, or hybrid
- Strength
- Broad vulnerability detection coverage with frequent plugin updates
- Watch for
- Noise from false positives requires tuning; licensing costs can escalate with add-ons
Qualys VMDR
- Pricing
- Starts at $0.25/asset/month; full VMDR from $2.50/asset/month
- Target
- Security teams needing integrated vulnerability management, detection, and response in a single cloud platform
- Deployment
- Cloud-based SaaS
- Strength
- Unified VM, threat detection, and patch management in one console
- Watch for
- Asset-based pricing can become expensive for large environments; complex initial setup
Rapid7 InsightVM
- Pricing
- Starts at $2,000/year for up to 256 assets; custom for larger deployments
- Target
- Security operations teams prioritizing real-time visibility and actionable risk prioritization
- Deployment
- On-prem or cloud-managed
- Strength
- Real-time dashboards and customizable risk scoring for quick remediation
- Watch for
- Higher cost per asset compared to some competitors; reporting can be complex to configure
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.