Axelera Europa AI Accelerator: 629 TOPS, 64GB per Chip and Shipping PCIe Systems
Axelera AI is now shipping Europa, its second-generation AI Processing Unit for edge and enterprise inference. The company announced the production ecosystem on September 15, 2026, with Europa available as a chip, a single-accelerator half-height PCIe card and a four-accelerator full-height server card. Validated systems include the Dell PowerEdge XE5 and Supermicro 111AD.
A Europa AIPU provides up to 629 TOPS at INT8, eight second-generation AI processing cores, 16 RISC-V vector-processing cores, 128MB of on-chip L2 SRAM and a 256-bit LPDDR5 interface rated at 200GB/s. Axelera lists support for up to 64GB of memory per chip and positions the architecture for generative AI, vision-language models, computer vision, robotics and other inference workloads.
The practical distinction is deployment density. Axelera lists single-chip configurations for models up to roughly 32 billion parameters and multi-chip cards for models up to roughly 100 billion parameters. These are vendor-stated capacity targets; actual model fit and throughput depend on model architecture, precision, context, KV-cache requirements and software support.
Europa specifications
| Specification | Europa |
|---|---|
| AI performance | Up to 629 TOPS at INT8 |
| AI cores | 8 second-generation AIPU cores |
| Vector processing | 16 RISC-V vector cores |
| Supported AI precision | INT4, INT8, INT16 |
| On-chip cache | 128MB L2 SRAM |
| DRAM interface | 256-bit LPDDR5 |
| DRAM bandwidth | 200GB/s |
| Memory capacity | Up to 64GB per chip |
| PCIe interface | PCIe Gen4 x4 |
| Package | 35 × 35mm |
| Published card TDP | 45W |
| Single-chip LLM capacity | Up to 32B parameters, vendor specification |
| Multi-chip LLM capacity | Up to 100B parameters, vendor specification |
| Software | Axelera Voyager SDK / Voyager Toolchain |
Europa uses Axelera's Digital In-Memory Compute architecture, which performs matrix operations inside SRAM arrays to reduce data movement. The new generation doubles the AI-core count relative to Metis and adds dedicated vector processors for pre- and post-processing. An integrated H.264/H.265 decoder can also move media decode work away from the host CPU for vision pipelines.
The PCIe cards and server options
Axelera is offering Europa across multiple integration levels. Customers can design boards around the AIPU itself, install an Edge 232p half-height, half-length PCIe card, or use the Server 250p full-height, full-length card for denser deployments.
The company says its PCIe portfolio scales from single-chip 16GB configurations to four-chip 256GB configurations. The September launch also moves Europa beyond an announced silicon architecture into available systems: Axelera says the Edge 232p is shipping and is available in validated Dell XE5 and Supermicro 111AD systems.
That availability distinction matters for infrastructure planning because Axelera originally announced the Europa AIPU architecture in October 2025. The September 2026 announcement covers the shipping product ecosystem and enterprise deployments rather than the first disclosure of the silicon itself.
Model capacity and inference performance
Axelera's current product material lists up to 32B parameters per Europa chip and up to 100B parameters across multiple chips. Memory capacity reaches 64GB per chip, giving Europa substantially more model capacity than the company's first-generation Metis platform.
For generative AI, Axelera currently publishes a 53.5 tokens/s result for Qwen3-30B-A3B on one Europa chip at a 1,024-token context, W8A8 precision and an 8-bit KV cache. The company explicitly labels this figure as a pre-silicon simulation, so it should be treated as a vendor projection until corresponding measured production-silicon results are published.
Axelera also claims 3–5× performance per dollar and 2–3× performance per watt over competing products in its category. Those are vendor comparisons whose usefulness depends on the exact workload and comparison hardware. The stronger independently verifiable specifications for deployment planning are the card formats, memory capacity, PCIe connectivity, TDP and supported software stack.
Software and model deployment
Europa uses the same Voyager software environment as Axelera's Metis accelerators. The common toolchain is intended to let developers move inference pipelines across embedded, edge and server hardware without adopting a separate programming environment for each product tier.
Axelera is also promoting AxeleraScript (AxScript) for model compilation and Voyager Wingman, an agentic development layer for porting and optimizing inference pipelines. Model compatibility remains a workload-specific consideration: teams evaluating Europa should verify that the exact model architecture and operators they require are supported by the current Voyager release before selecting hardware solely from memory capacity.
Where Europa fits
Europa targets inference rather than general-purpose GPU computing. Its combination of up to 64GB per chip, compact PCIe cards, 45W published card TDP and integrated pre/post-processing makes it particularly relevant to organizations deploying computer vision, multimodal models, robotics or local generative-AI services where rack power and data locality are important constraints.
The four-chip path is more interesting for larger models because it can provide substantially more aggregate memory than a conventional single accelerator card. Multi-chip model execution still depends on Axelera's software partitioning and the workload, so aggregate capacity alone is insufficient to predict latency or tokens per second.
For buyers comparing Europa with NVIDIA, AMD or other inference accelerators, the next useful evidence will be measured production-silicon benchmarks across common language and vision models, software coverage, street pricing and multi-chip scaling. Europa already has a clearer deployment story than at its 2025 architecture announcement because the September 2026 release adds shipping PCIe hardware and validated OEM systems.
Bottom line
Europa expands Axelera AI from its edge-focused Metis portfolio into higher-capacity enterprise inference. The core specifications are 629 INT8 TOPS, up to 64GB per chip, 200GB/s DRAM bandwidth, PCIe Gen4 x4 connectivity and a 45W published card TDP, with single- and four-chip PCIe configurations.
The most consequential September 2026 change is commercial availability: Europa is shipping in Axelera cards and validated Dell and Supermicro systems. Organizations considering it should base procurement on measured performance for their own supported models as production-silicon benchmark coverage expands.