Toradex Targets Edge AI with Its 32 TOPS-Capable Aquila AM69 System-on-Module
The first in a family of pin-compatible SOMs, the Aquila AM69 delivers a 3D-capable GPU, eight CPU cores, two MCUs, and a 32 TOPS NPU.
Swiss embedded computing specialist Toradex has announced the launch of the Aquila family of systems-on-modules (SOMs), offering support for both high-performance real-time computing and accelerated on-device machine learning and edge artificial intelligence (edge AI) workloads.
"With Aquila, we're thrilled to push the boundaries of embedded computing. We've leveraged extensive market research and invaluable customer feedback to develop this highly powerful yet cost-effective solution for high-performance applications," claims Toradex chief executive officer Samuel Imgrueth of the launch, in an announcement brought to our attention by Linux Gizmos. "Aquila represents our unwavering commitment to innovation and customer satisfaction, delivering unmatched performance, reliability, and value."
The company has confirmed just one model in the family at launch, the Aquila AM69 — powered, as the name suggests, by the Texas Instruments AM69A system-on-chip, giving it up to eight Arm Cortex-A72 application-class processor cores running at up to 2GHz and two Cortex-R5F real-time microcontroller cores running at up to 1GHz. To this, Toradex has added up to 32GB of LPDDR4 memory and up to 128GB of on-board eMMC storage.
It's the deep-learning accelerator that gives the module its claims to edge AI suitability, though: working alongside the main and microcontroller cores, the accelerator combines four C7X digital signal processor (DSP) cores with four matrix multiply accelerator (MMA) cores to deliver a claimed 32 tera-operations per second (TOPS) of compute — on top of what you can get out of the CPU cores and Imagination BXS 4-64 3D-capable graphics processing block.
According to Toradex's testing, the module — which exposes its capabilities over a 400-pin board-to-board interface that will be share with other boards in the family when they too launch — runs in a 30W thermal design profile (TDP). On the software side, the edge AI accelerator is supported in commonly-used frameworks including TensorFlow Lite and PyTorch, the company has confirmed, while a quad-lane MIPI Camera Serial Interface 2 (CSI-2) connection supports up to 12 Full HD camera inputs.
More information on the module is available on the Toradex website; at the time of writing, the company had yet to publicly announced pricing details.
Freelance journalist, technical author, hacker, tinkerer, erstwhile sysadmin. For hire: freelance@halfacree.co.uk.