Accelerate Edge AI Innovation
AI data-processing workloads at the edge are already transforming use cases and user experiences. The third-generation Ethos NPU helps meet the needs of future edge AI use cases.
The Ethos-U85 offers support for transformer-based models at the edge, the foundation for newer language and vision models, scales from 128 to 2048 MAC units, and is 20% more energy efficient than Arm Ethos-U55 and Arm Ethos-U65, enabling higher performance edge AI use cases in a sustainable way. Offering the same toolchain as previous Ethos-U generations, partners can benefit from seamless migration and leverage investments in Arm-based machine learning (ML) tools.
NPU
Overview
Key Features
- Sustainable Innovation
- 20% more energy efficient than Ethos-U55 and Ethos-U65, enabling future use cases in a sustainable way.
- Scalable Performance
- Scales from 128 to 2048 MACs, providing up to 4 TOPs of performance at 1 GHz.
- Generative AI at the Edge
- Native support for transformer networks, along with support for Tensor Operator Set Architecture (TOSA) as a standard.
- System Level Solution
- Supported by Arm Corstone-320, a reference design with a unified toolchain and the extensive Ethos-U ecosystem.
Benefits
- The Arm Ethos-U85 is the highest performance implementation of the Arm Ethos-U NPU. It enables enhanced edge AI processing with support for transformer-based models and delivers 20% more energy efficiency than previous generations of Ethos. The key characteristics of Ethos-U85 include:
- Scalable performance – 256 GOP/s up to 4 TOPs at 1 GHz
- Scales from 128 to 2048 MACs
- Further reduced energy consumption – 20% lower than previous Ethos-U NPUs
- Native support for transformer-based networks
- Scalable performance – 256 GOP/s up to 4 TOPs at 1 GHz
- Ethos-U85 targets numerous different applications with use in high-performance Arm Cortex-A or low-power embedded devices based on Arm Cortex-M.
Block Diagram
Technical Specifications
Related IPs
- AI accelerator (NPU) IP - 1 to 20 TOPS
- AI accelerator (NPU) IP - 16 to 32 TOPS
- AI accelerator (NPU) IP - 32 to 128 TOPS
- 4-/8-bit mixed-precision NPU IP
- NPU IP family for generative and classic AI with highest power efficiency, scalable and future proof
- AI Accelerator (NPU) IP - 3.2 GOPS for Audio Applications