Arm Ethos-N

Processor Series
NPU

Datasheet

AT A GLANCE

The Arm Ethos-N processor series delivers the

Powering AI Inference from Cloud to Edge
highest throughput and efficiency in the lowest area
for machine learning (ML) inference from cloud to
to Endpoint
edge to endpoint.
What’s New?

Network Support
Flexible design supports a variety of popular neural networks, including CNNs and
RNNs, for classification, object detection, image enhancements, speech recognition
and natural language understanding.

Futureproof Operator Coverage
The MAC engine flexibly decomposes arbitrarily sized kernels with stride and dilation
modes including convolution, deconvolution, depthwise separable, and vector
product. Programable Layer Engines execute layers not supported by the MAC engine,
supporting various primitives, activation functions and future operators.

Mixed Precision
Supports both Int-8 and Int-16: lower-precision Int-8 for classification and detection
tasks; high-precision Int-16 for HDR image enhancements and audio tasks.
Arm Ethos-N NPUs address the ML inference
requirements of multiple markets, providing 16,

Compression and Winograd Convolution
8 or 4 engines with identical software. MAC engines provide decompression, activation, Winograd transformation and scaling.
Winograd accelerates common filters by 225% compared to other NPUs, allowing
KEY FEATURES & BENEFITS actual performance to far exceed architectural performance.
S calable Performance 
Multicore
Delivering up to 4, 2 or 1 TOP/s of single
Supports up to eight processors in a tightly coupled cluster, with the ability to process
core performance with multicore scalability,
supporting up to eight NPUs in a cluster, multiple networks in parallel or a single, large network split across cores. Larger
and up to 64 NPUs in mesh systems. configurations of up to 64 cores are supported through Arm CoreLink mesh technology.
Highly Efficient
Achieving up to 5 TOPs/W (including 
Weight and Feature Map Compression
on-chip memory accesses) through Minimizes system memory bandwidth by 1.5-3x, reducing off-chip memory
internally distributed SRAM, storing data
accesses by 90% through extended compression technologies, targeting both weight
close to the compute elements to save
power and reduce DRAM access. and activations.
 Optimized Design

Security
Driving up to 225% convolution
performance uplift using Winograd Supports TrustZone system security with configurable secure queues for multiple users
on 3x3 kernels, delivering up to 90% and flexible processing in the TEE or SEE, providing layered security to protect both ML
MAC utilization.
models and input data.
 Futureproof
Supporting a wide range of existing ML 
System Integration (SMMU)
operations, as well as future innovations
ACE-Lite master port and optional SMMU (System Memory Management Unit)
through firmware updates and compiler
technology. integration allows for support and protection of memory and easy handling of multiple users..
KEY USE CASES FOR THE ETHOS Specifications
PROCESSOR SERIES

Object classification Ethos-N77 Ethos-N57 Ethos-N37

Key Features
Object detection Key Features Performance (at 1GHz) 4 TOP/s 2 TOP/s 1 TOP/s
Face detection/identification Mac/Cycle (8x8) 2048 1024 512
Human pose detection/hand-gesture
Data Types Int-8 and Int-16
recognition
Image segmentation Network Support CNN and RNN
Image beautification Efficient Convolution Winograd support
Super resolution
Sparsity Yes
Framerate adjustment
(super slow-mo) Secure Mode TEE or SEE
Speech recognition Multicore Capability 8 NPUs in a cluster
Sound recognition 64 NPUs in a mesh
Noise cancellation
Speech synthesis Memory System Embedded SRAM 1-4 MB 512 KB 512 KB
Language translation Bandwidth Reduction Extended compression technology,
layer/operator fusion, clustering,
and workload tilling
Main Interface 1xAXI4 (128-bit), ACE-5 Lite

Development Neural Frameworks TensorFlow, TensorFlow Lite,

Platform Caffe2, PyTorch, MXNet, ONNX
Neural Operator API Arm NN, AndroidNN
Software Components Arm NN, neural compiler, driver
and support library

Debug and Profile Layer-by-layer visibility

Evaluation and Arm Juno FPGA systems

Early Prototyping and cycle models

Market Segments

Mobile Smart camera STB/DTV Consumer

AR/VR Medical Robotics Drones

IoT Logistics Home Infrastructure

To find out more about the Ethos processor series, visit

developer.arm.com/ethos 10/19

All brand names or product names are the property of their respective holders. Neither the whole nor any part of the information contained in, or the product described in,
this document may be adapted or reproduced in any material form except with the prior written permission of the copyright holder. The product described in this document is
subject to continuous developments and improvements. All particulars of the product and its use contained in this document are given in good faith. All warranties implied or
expressed, including but not limited to implied warranties of satisfactory quality or fitness for purpose are excluded. This document is intended only to provide information to
the reader about the product. To the extent permitted by local laws Arm shall not be liable for any loss or damage arising from the use of any information in this document or
any error or omission in such information.

© Arm Ltd. 2019