New ARM IP Brings Down Costs of AI for Mainstream Products

October 28, 2019 by Gary Elinoff

Two new NPUs, along with a GPU and DPU, expand the scope of the doable for low-cost devices.

Applications such as AR (augmented reality), AI, and high-fidelity gaming are no longer the exclusive purview of premium devices. Increasingly, they are finding their way into the mainstream market.

Additionally, AI solutions—most importantly security—are no longer purely cloud-based. There is an increasing necessity for edge-based solutions, whether guarding against cyberattack or for faster, more realistic immersive reality experience.

As a part of this growing trend, Arm has just launched two new ML (machine language) processors, and has also augmented its popular Mali line with a new GPU and DPU pair. The new IP cores are aimed at bringing premium experiences to everyday, mainstream consumer devices.

The new suite of IP cores include:

  • Ethos-N57 and Ethos-N37 NPUs. These are aimed at AI applications, with an eye to balancing ML performance with cost, board space, and battery life. 
  • Mali-G57 GPU. This IP represents the first mainstream Valhall architecture-based GPU, making immersive gaming experiences possible for mainstream devices. 
  • Mali-D37 DPU. This delivers a rich display feature set and does so while requiring the smallest area within its host SoC. This digital processing unit is targeted at entry-level devices with small display screens.


Arm Ethos-N57 and Ethos-N37 NPU

The Ethos-N37 and Ethos-N57 are members of a family of NPUs (Network Processing Units) that includes the Ethos-N77. The use cases for the three devices are presented graphically in the illustration below.


Image from Arm


The Ethos-N77 is the Arm NPU designed for the most demanding applications. The new N37 and N57 are designed for SOCs serving mainstream applications with tighter area budgets. All members of this series are designed to minimize power-hungry DRAM traffic and to thereby conserve power usage—vital for any type of mobile device.

The capabilities of each of the devices are summarized as follows:


Device Throughput* MAC/Cycle** Internal Memory
Ethos-N37 Up to 1 TOP/s 512 8x8 512 KB
Ethos-N57 Up to 2 TOP/s 1024 8x8 512 KB
Ethos-N77 Up to 4 TOP/s 2048 8x8 1-4 MB

* TOP/s stands for terra operations per second

** Multiplier-Accumulator Operations


Mali-G57 GPU

The Mali-G57 GPU (graphics processing unit) aligns with the Vulkan API, and makes 4K/8K DTV user interfaces and console-like graphics on smartphones possible for mass-marketed devices.


The Mali-G57 Graphics Processor Unit. Image from Arm


This unit represents a 30% better energy efficiency which in turn translates into longer performance on a single battery charge. Foveated rendering support for VR concentrates image detail to where it is most needed, and designers can expect up to 60% performance improvement for ML workloads.


Mali-D37 DPU

The Mali-D37 is the first mainstream digital processing unit to be based on the Komeda architecture, which lays the groundwork significant savings in system power requirements. It is optimized for single displays with resolution of 1080p (full HD) to 2K. On 16mm technology, it is configurable to an area of less than 1mm.

This IP is aimed at inexpensive smartphones, tablets, and display applications including smart home appliances.



Do you work with machine learning in your job? Share your perspective on this technology in the comments below.