Arm Ethos-U85 NPU Brings Advanced AI to Edge

Arm has unveiled the Ethos-U85, its most powerful micro neural processing unit (NPU) to date, designed to run advanced AI inference workloads — including transformer models — on ultra-low-power embedded devices. The new chip delivers up to 4x the performance of its predecessor, the Ethos-U55, marking a significant leap in what edge devices can accomplish without cloud connectivity.

The Ethos-U85 arrives at a critical inflection point for the semiconductor industry, as demand for on-device AI processing surges across automotive, industrial IoT, wearables, and smart home applications. With this launch, Arm is positioning itself at the center of a rapidly expanding market that analysts project will exceed $50 billion by 2030.

Key Facts at a Glance

• Performance: Up to 4x improvement over the Ethos-U55 in ML inference throughput
• Power envelope: Designed for devices operating in the milliwatt range, targeting sub-1W total system power
• Transformer support: Native acceleration for transformer-based architectures, including small language models
• Scalability: Configurable from 128 to 2048 MAC (multiply-accumulate) units
• Ecosystem: Integrated into Arm's Corstone-320 reference design platform
• Toolchain: Full support through Vela compiler and compatibility with TensorFlow Lite for Microcontrollers

Why Edge AI Demands a New Class of NPU

The AI industry's center of gravity is shifting. While cloud-based large language models dominate headlines, a parallel revolution is unfolding at the edge — on the billions of microcontrollers embedded in everyday objects. These tiny processors increasingly need to run sophisticated AI workloads locally, without sending data to remote servers.

Privacy, latency, and bandwidth constraints are the primary drivers. A smart security camera that processes video locally eliminates the need to stream footage to the cloud. A predictive maintenance sensor in a factory must deliver real-time anomaly detection without network dependency.

Previous-generation NPUs like the Ethos-U55 handled basic convolutional neural networks (CNNs) effectively but struggled with more complex architectures. The rise of transformer models — originally designed for natural language processing but now dominating computer vision and audio tasks — created a capability gap that the Ethos-U85 directly addresses.

Ethos-U85 Architecture Delivers Major Performance Leap

The Ethos-U85 represents a ground-up architectural redesign rather than an incremental update. Arm has expanded the MAC array significantly, offering configurations that scale from 128 to 2048 MAC units. This flexibility allows silicon partners to tailor implementations for specific use cases and power budgets.

Key architectural improvements include:

• Enhanced memory subsystem: Larger internal SRAM buffers reduce external memory access, cutting both latency and power consumption
• Native transformer acceleration: Dedicated hardware paths for attention mechanisms and softmax operations
• Improved int8 and int16 support: Better quantization handling preserves model accuracy at lower precision
• Higher utilization rates: Architectural optimizations ensure the MAC array stays active more consistently during inference
• Backward compatibility: Models compiled for Ethos-U55 and Ethos-U65 can run on the U85 with recompilation

Compared to the Ethos-U65, which offered a mid-range option between the U55 and higher-end solutions, the U85 consolidates and surpasses both. Arm claims the new NPU achieves its performance targets while maintaining a comparable silicon area footprint, making it attractive for cost-sensitive embedded applications.

Corstone-320 Platform Accelerates Time to Market

Arm is not launching the Ethos-U85 in isolation. The NPU ships as part of the Corstone-320 reference design, which pairs it with the Cortex-M85 processor — Arm's highest-performance M-class CPU. This combination creates a complete subsystem that silicon vendors can integrate into their own chip designs.

The Corstone platform approach has proven effective in previous generations. It provides verified RTL, software development tools, and virtual hardware models that allow developers to begin writing and optimizing code before physical silicon becomes available. Arm estimates this approach can reduce development timelines by up to 18 months.

Virtual hardware is particularly significant. Through Arm's cloud-based simulation environment, developers can profile AI workloads, benchmark inference performance, and validate power consumption without access to physical development boards. This democratizes access for smaller teams and startups that may lack extensive hardware labs.

Transformer Models on Microcontrollers: A New Frontier

Perhaps the most consequential capability of the Ethos-U85 is its native support for transformer architectures. Until recently, running transformers on microcontroller-class devices was impractical. These models demand significant compute and memory resources that exceeded what embedded NPUs could deliver.

The Ethos-U85 changes this equation. Arm has demonstrated the NPU running small transformer models for tasks like keyword spotting, noise suppression, and visual anomaly detection — all within milliwatt power budgets. While these are far from the billion-parameter models running in data centers, they represent a meaningful step toward more intelligent edge devices.

This capability opens several practical applications:

• Voice interfaces: More natural and accurate voice command recognition on smart home devices
• Predictive maintenance: Advanced anomaly detection in industrial equipment using time-series transformers
• Health monitoring: Wearable devices that can run more sophisticated biosignal analysis locally
• Smart agriculture: Sensor nodes capable of complex environmental pattern recognition
• Automotive: In-cabin monitoring and gesture recognition without dedicated high-power processors

The industry trend toward TinyML — machine learning optimized for microcontrollers — aligns perfectly with the Ethos-U85's capabilities. Organizations like the MLCommons consortium have established benchmarks specifically for this class of device, and the U85 is expected to set new performance records in upcoming benchmark submissions.

Competitive Landscape Heats Up in Edge AI Silicon

Arm's announcement does not happen in a vacuum. The edge AI silicon market has grown increasingly competitive, with multiple players vying for design wins in the embedded space.

Qualcomm continues to push its Hexagon NPU technology into lower-power tiers. STMicroelectronics has integrated neural processing capabilities into its STM32 microcontroller line. Intel (through its acquisition of Habana Labs and Movidius) and Google (with its Edge TPU) also compete for edge AI workloads, though typically at higher power points.

Startups like Syntiant, Hailo, and Kneron have carved out niches with specialized ultra-low-power AI chips. Syntiant's NDP200, for example, targets always-on audio processing at microwatt-level power consumption — a segment even below the Ethos-U85's target range.

Arm's key advantage remains its ecosystem. With over 250 billion Arm-based chips shipped to date and a vast network of silicon partners — including NXP, Samsung, Renesas, and Ambiq — the company's NPU technology benefits from unmatched distribution reach. When Arm introduces a new NPU, it potentially influences chip designs across dozens of semiconductor companies.

What This Means for Developers and Businesses

For embedded developers, the Ethos-U85 lowers the barrier to deploying sophisticated AI models on resource-constrained devices. The improved Vela compiler toolchain handles model optimization and quantization, translating standard TensorFlow Lite models into efficient NPU instructions with minimal manual intervention.

Businesses evaluating edge AI strategies should note several implications. First, the performance headroom means that AI features previously requiring application-processor-class hardware (think Cortex-A series) can now potentially run on cheaper, lower-power microcontroller platforms. This could reduce bill-of-materials costs for IoT products by $2 to $10 per unit — significant at scale.

Second, the privacy benefits of local AI processing are increasingly important as regulations like the EU's AI Act and GDPR impose stricter requirements on data handling. Devices that process sensitive data locally without cloud transmission simplify compliance.

Third, the reduced dependency on network connectivity makes edge AI solutions viable in environments where connectivity is unreliable or unavailable — remote industrial sites, agricultural deployments, and developing markets with limited infrastructure.

Looking Ahead: The Road to Ubiquitous Edge Intelligence

Arm's Ethos-U85 represents a clear trend toward making AI inference a standard capability in even the smallest and most power-constrained computing devices. As model compression techniques improve and NPU architectures advance, the gap between cloud AI and edge AI capabilities will continue to narrow.

Several developments to watch in the coming 12 to 18 months include the first commercial chips incorporating the Ethos-U85 from Arm's silicon partners, expected in late 2025 or early 2026. Benchmark results from MLCommons' MLPerf Tiny suite will provide independent validation of the NPU's performance claims.

The broader trajectory points toward a future where billions of devices — from industrial sensors to consumer electronics — possess meaningful AI capabilities without requiring cloud connectivity or significant power budgets. Arm's strategy with the Ethos-U85 is to ensure its architecture sits at the heart of that transformation, embedded in chips from dozens of semiconductor vendors worldwide.

For an industry that has spent the past 2 years focused almost exclusively on data center AI, the Ethos-U85 is a timely reminder that some of AI's most transformative applications will run not in massive server farms, but on chips smaller than a fingernail.