Edge AI – How AI is sparking the adoption of edge computing

Edge Insights

Purchase report

This report is available to purchase.

Buy Now

Login to access

Want to subscribe?

This article is part of: Edge Insights

To find out more about how to join or access this report please contact us

AI applications will require low-latency, local compute for rapid inferencing and large scale data collection, triage, and engineering. Edge compute will therefore play a key role in AI app delivery. However it's not just about infrastructure - commercial scale for edge AI will depend on effective ecosystem collaboration models.

Enterprises across all industries are investigating how they can leverage AI applications

Since the release of GPT-3 from OpenAI, consumer interest in AI has exploded. This change in consumer behaviour follows that of enterprises, who have been investigating AI and its application to their processes for years.

AI has been around for decades. The recent revolution in AI capabilities stems from a combination of innovation in hardware technology and the deep learning era; beginning in 2010, a subset of machine learning used multi-layered neural networks to analyse various forms of data. Since 2012, companies creating these deep learning models have seen an increase in their financing, allowing them to build large data centres with the specific hardware (GPUs) to process larger datasets.

The graph below illustrates this increase in compute power usage by machine learning systems over time.

Trends in training of machine learning and AI models

Source: “Compute trends across three eras of machine learning”, 2022, J. Sevilla et al. https://arxiv.org/pdf/2202.05924.pdf

Large Language Models (LLMs) like Chat GPT are the most widely known AI models. Training these models on billions of tokens sourced from the public internet (GPT-3: 175 billion parameters, LaMDA [Google]: 137 billion parameters) allows these LLMs to creatively solve a wide variety of problems from equally varied inputs. This ability to ingest diverse inputs and produce outputs which can be utilised in a wide range of environments, is the key differential that has sparked the wave of recent interest.

Horizontal adoption of AI models has exploded in the past 18 months

Source: Wikibon

Enter your details below to download an extract of the report

Enterprises will access centralised AI models, developed and trained in the cloud

AI is not consigned to just language processing. Although LLMs have been the main driver of interest since 2022, AI models such as convolutional neural networks (CNNs) and graph neural networks (GNNs) can ingest different data structures (image/video and graphical, respectively) to provide valuable insights.

Most enterprises are still exploring the possibilities of AI. Initial implementations are centred around simple tasks, such as administration or customer service, but executives are exerting pressure on their organisations to explore how AI can automatically manage mission-critical tasks and augment productivity. To do this effectively and reliably, enterprises will need AI models that are well designed and well trained to carry out the mission-critical tasks that are required of them.

Few enterprises will have the resources to create and train their own AI model for a specific application within a specific industry. This would require a huge amount of compute power, which is both expensive and scarce, as well as a large amount of well documented and parsed data. Bloomberg has created its own LLM, BloombergGPT, pooling its extensive archive of financial data (363 billion tokens) with a public dataset (345 billion tokens) to train the model. Companies like OpenAI and Inflection.AI continue to raise huge sums of capital investment to build their own. However, few enterprises will have the resources to follow suit.

Most will leverage models trained on large, centralised datasets (like GPT) that are built in the cloud, fine tuning them through prompts and contextualisation to create outputs which are more tailored for their environment. This approach is considered less secure than building your own, as it involves the sharing of proprietary data with the model, which will then use these inputs to further train its algorithm. However, this approach allows enterprises to benefit from the generalist applications of these models whilst not having to invest time and money in developing and training the model themselves.

Table of contents

  • Executive Summary
    • The role of edge AI in driving scale for AI-driven applications
    • Effective ecosystem management will underpin the success of AI at the edge
  • Introduction
    • Enterprises across all industries are investigating how they can leverage AI applications
    • Enterprises will access centralised AI models, developed and trained in the cloud
  • Interest in AI is driving demand for edge but obstacles remain
    • Technical challenges
    • Commercial challenges
    • Enterprises are concerned about the security of AI and protecting their proprietary data
  • An ecosystem approach to edge AI
    • What is an ecosystem?
    • Why are ecosystem approaches necessary for scaling edge AI?
  • Conclusion

Related research

Enter your details below to download an extract of the report