AI Hardware - Edge Machine Learning Inference


AI Hardware - Edge Machine Learning Inference

With the growing demand for real-time deep learning workloads, today’s standard cloud-based Artificial Intelligence approach is not enough to cover bandwidth, data privacy or low latency applications. Hence, technology is required to run AI without the cloud, moving AI to the edge This drives the need for specific AI hardware for on-device machine learning inference.

Computer vision and artificial intelligence are transforming IoT devices at the edge. In this article you will learn about specialized AI hardware, also called AI accelerators, created to accelerate data-intensive deep learning inference on edge devices in a cost-effective way. Particularly, you will learn:
1. Machine learning inference (Basics)
2. The need for specialized AI hardware
3. List of the most popular AI accelerators in 2021

Machine Learning Inference at the Edge

Inference is the process of taking a neural network model generally made with deep learning, and then deploying it onto a computing device. This device will then process incoming data (usually images or video) to look for and identify whatever it has been trained to recognize.

While deep learning inference can be carried out in the cloud, the need for edge inference is growing rapidly due to bandwidth, privacy concerns or the need for real-time processing.

Installing a low power computer with an integrated AI inference accelerator, close to the source of data, results in much faster response times and more efficient computation. It requires less internet bandwidth and graphics power. When compared to cloud inference, inference at the edge can potentially reduce the time for a result from a few seconds to a fraction of a second.

People Detection with Edge AI Inference

The Need for Specialized AI Hardware

Today, enterprises are extending analytics and business intelligence closer to the points where data is generated. Edge computing solutions place the computing infrastructure closer to the source of incoming data. This also places them closer to the systems and people who need to make data-driven decisions in real-time. In short, the AI model is trained in the cloud and deployed on the edge device.

Especially in computer vision the workloads are high and tasks to be computed are highly data-intensive. AI hardware acceleration for edge devices has many advantages, the main being speed:
1. Speed and performance. By processing data closer to the source, edge computing greatly reduces latency. The end result is higher speeds, enabling real-time use cases.
2. Better security practices. Critical data does not need to be transmitted across different systems. User access to the edge device can be very restricted.
3. Scalability. Edge devices are endpoints of an AI system that can grow without performance limitations. This allows to start small and with minimal costs. The development of cloud-based technology and edge computing have made it easier than ever for businesses to scale their operations.
4. Reliability. Edge computing distributes processing, storage and applications across a wide range of devices, making it difficult for any single disruption to take down the network (cyberattacks, DDoS attacks, power outages, etc.).
5. Offline-Capabilities. An Edge-based system is able to
work despite limited network connectivity, a crucial factor for mission-critical systems.
6. Better data management. Fewer bottlenecks through distributed management of edge nodes. Only processed data of high-quality is sent to the cloud.
7. Privacy. Sensitive data can be processed locally and in real-time, without streaming it to the cloud.
AI accelerators can greatly increase the inference or execution speed of an AI model, and can also be used to execute special AI-based tasks that cannot be conducted on a CPU.

Most Popular Edge AI Hardware Accelerators

With AI becoming a key driver of edge computing, the combination of hardware accelerators and software platforms are becoming important to run the models for inferencing. NVIDIA Jetson, Intel Movidius Myriad X or Google Coral Edge TPU are popular options available to accelerate AI at the edge.
1.) VPU: Vision Processing Unit
VPUs allow demanding computer vision and edge computing AI workloads to be conducted with high efficiency. VPUs achieve a balance of power efficiency and compute performance.
One of the most popular examples of a VPU is the Intel Neural Computing Stick 2 (NCS 2), that is based on the Intel Movidius Myriad X VPU By running programmable computation strategies in parallel with workload-specific hardware acceleration, Movidius Myriad X creates an architectural environment that minimizes data movement.
The Intel Movidius Myriad X VPU is Intel’s first VPU that features the Neural Compute Engine – a highly intelligent hardware accelerator for deep neural network inference.
The Intel Movidius Myriad X VPU is programmable with the Intel Distribution of the OpenVINO Toolkit Used in conjunction with the Myriad Development Kit (MDK), custom vision, imaging, and deep neural network workloads can be implemented using preloaded development tools, neural network frameworks, and APIs.

2.) GPU: Graphics Processing Unit
A GPU is a specialized chip that can do rapid processing, particularly handling computer graphics and image processing. One example of devices bringing an accelerated AI performance to the Edge in a power-efficient and compact form factor is the NVIDIA Jetson device family.
The NVIDIA Jetson Nano development board, for example, allows neural networks to run using the NVIDIA Jetpack SDK. In addition to a 128-core GPU and Quad-core ARM CPU, it comes with nano-optimized Keras and Tensorflow libraries, allowing most neural network backends and frameworks to run smoothly and with little setup.

3.) TPU: Tensor Processing Unit
A TPU is a specialized AI hardware that implements all the necessary control and logic to execute machine learning algorithms, typically by operating on predictive models such as artificial neural networks (ANN)
The Google Coral Edge TPU is Google’s purpose-built ASIC designed to run AI at the edge. The Google Coral TPU is a toolkit built for Edge that enables production with local AI. More specifically, the onboard device inference capabilities of Google Coral TPU allow users to build and power a wide range of on-device AI applications. Core advantages are the very low power-consumption, cost-efficiency and offline-capabilities.
Google Coral devices are able to run machine learning frameworks (such as TensorFlow Lite, YOLO, R-CNN, etc.) for Object Detection to detect objects in video streams from connected cameras.

Google Coral AI Accelerator TPU – USB version (Source: Google Coral)