Power-efficient edge AI processing targets next-generation HMI

members can Download this article in PDF format.

We’ve been immersed in traditional human-machine interface (HMI) methods for so long that the addition of artificial intelligence (AI) and edge computing requires an optimization of our knowledge base. In the past, an HMI consisted of a physical control panel with pushbuttons, switches, trackballs, and indicator lights that allowed users to communicate with a machine. But it’s not sacrilegious to say they’re leaving the scene quickly.

In their place, edge computing brings computing power to where your data is collected. Edge AI is more nimble and agile than other forms of computing that send data to remote data centers or the cloud for processing. It offers faster, localized processing with lower latency than traditional forms of cloud computing.

With the proliferation of AI and machine learning (ML), the next generation of HMI is poised to open up entirely new ways of interacting with machines and devices. Pairing edge AI capabilities like machine vision, analytics, and predictive maintenance with HMI applications is changing the way we interact with machines. Examples are gesture and/or face recognition in a noisy factory environment, or controlling machines via digital voice assistants (like Alexa and Siri) and a wireless connection.

Sponsored Resources:

Taking HMI to the next level

As the HMI continues to evolve, the processor technology behind these applications must be ready to accommodate this evolution. Recently, Texas Instruments released new Edge AI processors for HMI applications. The first devices in TI’s Sitara AM62 processor family, the AM625 and AM623 are designed for low power consumption and bring low-power edge AI processing to dual-display and small-form applications. It includes an AI accelerator that achieves high deep learning AI inference performance. Inference involves the process of AI analyzing data collected from a sensor and making decisions based on that data in real time or near real time.

TI has put a lot of effort into developing the AM62x family (Fig. 1). The processors consist of a quad-core 64-bit Arm Cortex-A53 microprocessor, a single-core Cortex-R5F microcontroller (MCU), and a Cortex-M4F MCU. While the AM623 targets IoT and gateway SoCs with Arm Cortex-A53-based object and gesture recognition, the AM625 specifically targets Arm Cortex-A53-based edge AI and Full HD dual-display capability Human-Machine Interactions from .

Perhaps the best part is the low power consumption of these devices. AM62x processors achieve less than 500mW power consumption for a single A53 core running at 1GHz. That’s almost half of what similar low-power, low-cost processors of previous generations achieved. AM62x processor power can be less than 1W when quad A53 cores run a strenuous application on all four cores at 1.4GHz.

To bring low-power analysis to edge devices, including suspend states as low as 7mW, without thermal considerations, TI engineers applied some clever techniques.

Low power design

The low power consumption was made possible by a simplified power delivery architecture – the device has only two dedicated power rails and five power delivery modes. This device’s modular architecture delivers performance with support for multiple power saving modes without sacrificing critical system resources such as connectivity, performance, security, and cost.

Deep sleep mode at <5mW enables longer battery life, while active power of <1.5W is enabled by a core voltage of 0.75V. This allows AM62 processors to reduce power consumption in industrial applications by up to 50% compared to competing devices, allowing an application powered by AA batteries to remain on for over 1,000 hours.

Achieving optimal performance is further simplified with the new TPS65219, a companion power management IC (PMIC) designed specifically for the power supply requirements of AM62x processors. The device is characterized by an ambient temperature range of -40 to +105 °C, which makes it suitable for various industrial applications. It includes three synchronous step-down DC/DC converters and four linear regulators.

Additional functions

The AM62x processors include on-chip resources such as a general purpose asynchronous receiver/transmitter, multi-channel SPI and I²C, as well as various connection options for common industrial sensors or controls. These processors also offer dual Ethernet support and EtherCAT master support with a third-party ecosystem.

The parts also support a variety of display interfaces, including RGB888 color format, which uses 8 bits for each of the red, green, and blue color components, and a low-voltage differential signaling interface that supports 2K and full high-definition displays. The processors support dual displays, which offer design flexibility to engineers as two screens effectively expand the visual realm and help make the workflow smoother.

In addition, AM62x processors support multiple operating systems including mainline Linux and Android operating systems.

The peripherals included in the AM62x enable system level connectivity such as USB, MMC/SD cards, camera interface, CAN-FD and GPMC (General Purpose Storage Controller) for a parallel host interface to an external ASIC/FPGA. The AM62x device also supports secure boot for IP protection with an integrated hardware security module (HSM) and advanced power management support for portable and power-sensitive applications.

The AM625 and AM623 processors are available in a 13mm x 13mm, 425-pin ALW package and start at less than $5 in 1,000-unit quantities. The processors can meet the AEC-Q100 automotive standard in a 17.2 × 17.2 mm package (AMC). Industrial and automotive functional safety requirements are met using the integrated Cortex-M4F cores and dedicated peripherals, all of which can be isolated from the rest of the AM62x processor.