What Is INT8 Quantization?

INT8 quantization converts an AI model’s weights and calculations from 32-bit floating point (FP32) to 8-bit integers. The model becomes about a quarter of the size and runs faster on NPUs, GPUs and AI accelerators, usually with only a small loss of accuracy. It is a standard step when preparing a trained model for inference at the edge.

GPU vs CPU: What Is the Difference?

A CPU has a small number of powerful cores that handle many different tasks one after another, such as running the operating system and applications. A GPU has thousands of simpler cores that work in parallel, which makes it far faster for graphics, video and AI inference, where the same calculation is repeated across large amounts of data.

What Are the COM Express Module Types?

COM Express is a PICMG standard for computer-on-modules, and its pin-out types define which interfaces a module provides. Type 6 is the most common, with display outputs, USB and PCIe for general embedded systems. Type 7 is headless, with up to four 10GbE ports for edge server use. Type 10 is a smaller, lower-power pin-out for Mini modules.

OpenVINO is Intel’s open-source toolkit for optimising and running AI inference. It takes models trained in frameworks such as PyTorch, TensorFlow or ONNX, optimises them, and runs them efficiently on Intel hardware, including CPUs, integrated GPUs such as Iris Xe and Arc, discrete Arc GPUs and the NPUs in Core Ultra processors.

iGPU vs dGPU: What Is the Difference?

An iGPU (integrated GPU) is built into the processor and shares the system memory, so it adds little power, heat or cost. A dGPU (discrete GPU) is a separate graphics processor with its own memory, offering far more performance but needing more power, cooling and space. For edge AI, iGPUs suit lighter inference and dGPUs suit heavy workloads.

What Is a Reverse Vending Machine?

A reverse vending machine (RVM) takes back empty drinks containers, such as plastic bottles, cans and glass, and gives the user a refund or reward. It identifies each container by its barcode and shape, checks it is eligible, then crushes, sorts and stores it. RVMs are the collection point for deposit return schemes.

AI Inference vs Training: What Is the Difference?

Training is how an AI model learns: it is fed large sets of labelled data and adjusted until it predicts accurately, which needs powerful GPUs. Inference is using the trained model on new data to make a decision, which is lighter per image and usually runs in real time on an edge computer beside the camera or machine.

COM Express SBC Utilises Intel Iris Xe Graphics for Deep Learning AI

TypeCase study

Published5 May 2023

AuthorAndrew WhitehouseManaging Director

Reading time6 min

Reverse vending machines (RVMs) reward people for recycling: insert a plastic bottle, glass bottle or can, and the machine scans, sorts and stores it, then pays out cash or a voucher. As deposit return schemes spread, machines have to recognise containers faster and more reliably, and that is increasingly a job for AI.

An RVM manufacturer integrated DFI’s TGU968 COM Express module into its latest machine, using the 11th Gen Intel Core processor and its integrated Iris Xe graphics to run deep learning recognition. It is the kind of system we supply for retail and self-service.

How it works · 4 steps Reverse vending

Inside a reverse vending machine

Every container goes through the same sequence in seconds, and the AI step decides how fast the queue moves.

Insert
Container in

The user feeds in a bottle or can, and the machine moves it past its cameras and scanners.
Recognise
Deep learning

Machine vision and AI identify the container by its shape, material and label, checking it is eligible for a refund.
Sort
Crush and store

Approved containers are crushed or kept whole and sent to the right storage bin.
Reward
Refund

The machine issues cash, a voucher or a digital reward.

Module · 4 points

Why a COM Express module

01Module plus carrier: The module carries the processor and memory and plugs into a carrier board built for the machine’s I/O.
02Upgradeable: A newer module of the same type can replace the old one, so the machine’s compute can improve without redesigning its electronics.
03Simpler architecture: One standard compute block across the product range simplifies design and long-term support.
04Compact form factor: A Compact module keeps the electronics small inside a machine full of mechanics, cameras and sensors; see our COM Express module types FAQ.

Our system-on-module explainer covers how the module and carrier split works, and our COM-HPC vs COM Express comparison shows when the newer standard is worth the move. We also supply COM Express carrier boards.

Products · 1Module

The DFI TGU968

An 11th Gen Intel Core COM Express Compact module with Iris Xe graphics and 2.5GbE.

DFI

TGU968

11th Gen Intel COM Express Compact Module with 2.5GbE

Processor: Intel Core 11th Gen (Tiger Lake)
Networking: 1× 2.5GbE
Memory: 2× DDR4-3200 SO-DIMM, up to 64 GB

View product

AI · 4 points

Deep learning on integrated graphics

01Iris Xe on the processor: The Tiger Lake Core i5 and i7 processors include Iris Xe graphics on the same chip, sharing system memory instead of needing a separate GPU.
02Lower size, power and cost: No discrete graphics card means a smaller, cooler and cheaper system, removing a traditional barrier to AI in machines.
03Optimised with OpenVINO: OpenVINO optimises trained models to run efficiently on the CPU and Iris Xe graphics together.
04Built for inference: Iris Xe handles inference, running trained models on new images; training stays on more powerful hardware.

For heavier workloads, such as many camera streams or training large models, a discrete GPU or a dedicated AI platform is the better fit; our iGPU vs dGPU FAQ sets out the trade-offs, and our inference vs training FAQ explains the difference between the two workloads.

Intel Xe graphics combined with the OpenVINO inference engine reach performance levels previously only seen on much more expensive AI platforms. It opens AI up to a wider audience, giving software developers and ML professionals the tools to build applications that make a difference.

Andrew Whitehouse Managing Director, Things Embedded

Results · 4 outcomesReverse vending

What the module delivered

Integrated graphics and OpenVINO brought deep learning into the machine without adding a GPU.

Outcome 01

Up to five times faster

Recycling speed increased up to five times compared with traditional single-feed machines.
Outcome 02

Less waiting, more recycling

Faster recognition cuts the time each return takes, improving the effectiveness of the process.
Outcome 03

AI without a discrete GPU

Iris Xe and OpenVINO kept size, power and cost down.
Outcome 04

Ready to upgrade

The COM Express design lets future modules drop in as the technology improves.

06FAQs

COM Express and edge AI FAQs

What Is INT8 Quantization?

INT8 quantization converts an AI model’s weights and calculations from 32-bit floating point (FP32) to 8-bit integers. The model becomes about a quarter of the size and runs faster on NPUs, GPUs and AI accelerators, usually with only a small loss of accuracy. It is a standard step when preparing a trained model for inference at the edge.
GPU vs CPU: What Is the Difference?

A CPU has a small number of powerful cores that handle many different tasks one after another, such as running the operating system and applications. A GPU has thousands of simpler cores that work in parallel, which makes it far faster for graphics, video and AI inference, where the same calculation is repeated across large amounts of data.
What Are the COM Express Module Types?

COM Express is a PICMG standard for computer-on-modules, and its pin-out types define which interfaces a module provides. Type 6 is the most common, with display outputs, USB and PCIe for general embedded systems. Type 7 is headless, with up to four 10GbE ports for edge server use. Type 10 is a smaller, lower-power pin-out for Mini modules.
What Is OpenVINO?

OpenVINO is Intel’s open-source toolkit for optimising and running AI inference. It takes models trained in frameworks such as PyTorch, TensorFlow or ONNX, optimises them, and runs them efficiently on Intel hardware, including CPUs, integrated GPUs such as Iris Xe and Arc, discrete Arc GPUs and the NPUs in Core Ultra processors.
iGPU vs dGPU: What Is the Difference?

An iGPU (integrated GPU) is built into the processor and shares the system memory, so it adds little power, heat or cost. A dGPU (discrete GPU) is a separate graphics processor with its own memory, offering far more performance but needing more power, cooling and space. For edge AI, iGPUs suit lighter inference and dGPUs suit heavy workloads.
What Is a Reverse Vending Machine?

A reverse vending machine (RVM) takes back empty drinks containers, such as plastic bottles, cans and glass, and gives the user a refund or reward. It identifies each container by its barcode and shape, checks it is eligible, then crushes, sorts and stores it. RVMs are the collection point for deposit return schemes.
AI Inference vs Training: What Is the Difference?

Training is how an AI model learns: it is fed large sets of labelled data and adjusted until it predicts accurately, which needs powerful GPUs. Inference is using the trained model on new data to make a decision, which is lighter per image and usually runs in real time on an edge computer beside the camera or machine.