Axelera AI has released Europa, a new system-on-chip designed to lower the cost of running large language models. This launch matters because it targets the expensive inference workloads that currently dominate data center budgets. Buyers looking for specialized hardware to reduce operational costs now have a new option to evaluate.
New AIPU targets lower inference costs with dedicated hardware
The Europa chip is an AI Processing Unit (AIPU) built for high-efficiency inference. It replaces general-purpose graphics processors with a dedicated architecture optimized for specific AI tasks. This approach aims to deliver higher performance per dollar for developers deploying language models.
Specifications
- AI Compute Performance: 629 TOPS
- AI Processing Cores: 8
- Vector Processing Cores: 16 RISC-V
- On-chip L2 Cache: 128MB
- Power Consumption: 45W
Europa delivers 629 TOPS of AI compute power using eight dedicated AI processing cores. The chip integrates 16 RISC-V vector processing cores and 128MB of on-chip L2 cache to manage data flow efficiently. It supports INT4, INT8, and INT16 data formats and connects via a 256-bit LPDDR5 memory interface.
The system includes four PCIe Gen4 lanes for connectivity and maintains a total power consumption of just 45W. Axelera AI claims this efficiency allows Europa to achieve up to six times the token throughput per watt compared to standard GPU solutions. At equivalent price points, the company states the chip offers more than eight times the throughput of its best competitors.
Axelera AI officially launched this second-generation AIPU chip on September 15, 2024. The company has not yet disclosed specific pricing or availability dates for the Europa chip. The hardware is positioned as a cost-effective alternative to traditional GPU clusters for AI inference tasks.



Discussion
0 comments
Log in to join the thread with a thoughtful take, question, or correction.