Embedded Technology
AI needs more than one size to fit all requirements
By Martin Kellermann, marketing manager, Microchip Technology A
pplications of artificial intelligence (AI) are appearing everywhere. Widespread adoption of cloud-based tools and interactive services show
what the technology can do for the virtual world. But the physical world is where much of the capability has yet to be exploited. There are many examples of what is possible. AI models in sensor hubs attached to machine tools and other equipment can detect when wear on the components is degrading performance. Armed with cameras, robotic lawnmowers can tend the garden and avoid hurting wildlife and pets. Complex arrays of sensors and cameras can keep watch over busy areas, making sure they stay safe. Other systems take control of increasingly complex robots; these range from self-driving industrial vehicles to humanoid designs that can handle intricate assembly and repair jobs. Each of these applications has different demands in terms of the types of AI model they need to support their core functions. Though language models seem to dominate the cloud environment now, the landscape for local AI is far more diversified. These systems also need local processing at the edge to overcome the problems of latency and communication cost that come with keeping AI execution in the cloud.
Some applications call for the use of convolutional neural networks (CNNs). Others, such as those that interpret sensor data, make use of the recurrent neural network (RNN) structure. Systems that interpret complex scenes captured by cameras and other 2D sensors may use the capabilities offered by vision-language models or other variants of foundation models.
You cannot use a one-size-fits-all approach to hardware design. These disparate model architectures have very different computational requirements. A good analogy is the range of options for transportation: from lightweight bicycles to heavy trucks. For certain tasks, an e-bike is a good choice
38 July/August 2026
because it is highly efficient over short distances with light packages, such as food deliveries. But for bulk transport, you need the power and capacity of a 40-tonne truck. Similarly with AI, there are solutions that best fit the task at hand. They range from microcontrollers to heavyweight systems based on multiple AI accelerators. As with e-bikes, cars, vans and trucks, there are many options within each class of vehicle for engine performance and carrying capacity. Hardware for AI needs similar tuning capabilities: performance relies not just on arithmetic throughput but on the efficiency of data movement and organisation. This is where the kinds of hardware acceleration made possible by microprocessor-based acceleration and the field-programmable gate array (FPGA) come into their own. Often, model quality improves with preprocessing of the incoming data, such as mapping full-colour pixels into separate red, green and blue planes. They can benefit from other low-level manipulations of the data, such as normalisation or conversion
Components in Electronics
into different spaces. The models based on transformers benefit from hardware that can fetch data from across the memory space and pipeline the elements into the arithmetic arrays.
Models with lower computational needs, such as CNNs, can often be optimised by techniques such as pruning and data quantisation. Most teams train models using floating-point arithmetic on cloud servers. This maximises flexibility and streamlines the process of training.
During inference, there is often no need to process data with the high resolution that floating-point supports. Conversion of weights and input data to 8-bit integers, and sometimes even lower resolutions, will frequently deliver only slightly reduced accuracy.
Pruning removes links between neurons that have little effect on the network’s output. As most AI accelerators are designed for dense networks, some reorganisation of the data in memory will be needed to enable high throughput. FPGA-based
address generators can handle these tasks easily. These can stream data spread across the memory map into an execution pipeline designed for dense networks. Other techniques, such as Flash attention for transformers, also use intelligent memory mapping and accesses to deliver weights and data in the most efficient order. Several examples will show both the range and value that today’s hardware platforms can provide. One of the key applications of AI lies in image and video processing. We can think of this as the 3-tonne van of the edge-AI space: a versatile vehicle for a wide range of tasks. Industrial and commercial organisations can deploy visual models to take care of monitoring for safety, security and efficiency. Autonomous lawnmowers can combine a model trained to distinguish plants and objects to perform local planning and calculate the most- efficient path. In the industrial environment, vision-based control can detect different types of source material placed in an oven, whether foodstuffs or structural composites,
www.cieonline.co.uk
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48
