Axelera AI Europa

Overview

The Axelera AI Europa is an enterprise-grade artificial intelligence accelerator fabricated on Samsung’s advanced 5nm process node, engineered to deliver high-density neural processing across edge and data center infrastructures. At the architectural core, Europa combines 8 dedicated AI processing cores with 16 integrated RISC-V vector cores, allowing the chip to handle intensive neural network inference alongside complex pre- and post-processing routines directly on-die. Memory performance is prioritized through an expansive 128MB on-chip L2 SRAM cache coupled to a 256-bit LPDDR5 memory interface, pushing up to 200 GB/s of local memory bandwidth and supporting configurations up to 64GB per chip. An integrated on-chip hardware video decoder further minimizes bus congestion by ingesting and processing high-throughput video streams without routing uncompressed frames over system PCIe buses.

From a thermal and physical deployment standpoint, the accelerator is designed to fit flexible enterprise form factors including bare silicon, the half-height half-length (HHHL) Axelera Edge 232p, and the full-height full-length (FHFL) Axelera Server 250p. Operating within a tightly bounded 45W Thermal Design Power (TDP) envelope, Europa achieves an efficiency metric of up to 629 TOPS across INT4, INT8, and INT16 precisions. This modest thermal dissipation enables broad compatibility with standard server chassis airflow solutions without requiring high-maintenance liquid loops or complex custom thermal assemblies. Physical reliability and server validation are reinforced through qualification in tier-1 enterprise platforms such as the Dell PowerEdge XE5 and Supermicro 111AD servers, although broad enterprise adoption remains linked to continuing driver and model runtime optimization across competitive open-source machine learning stacks.

Technical Specifications