Brandon Wang is vice president of Synopsys.
The rapid development of AI has led to significant growth across the computing industry. But it is also causing a huge increase in energy consumption, which is leading us into an energy crisis. Current AI models, especially large language models (LLMs), need huge amounts of power to train and run. AI queries require much more energy than traditional searches; for example, asking ChatGPT a question consumes up to 25 times as much energy as a Google search. At current rates of growth, AI is expected to account for up to 3.5% of global electricity demand by 2030, twice as much as the country of France.
We need to address this issue urgently before it becomes unsustainable. If we don’t, the impact could threaten sustainable growth and the widespread adoption of AI technologies themselves. Fortunately, there are a number of pathways toward more energy-efficient AI systems and computing architectures.
Hardware And Software Solutions
On the hardware side, the “shift-left” approach means redesigning computer architectures with energy consumption in mind from early in the design process, optimizing power savings at every point in the design phase. This leads to better overall performance and cost savings through more power-efficient hardware. Fixing issues during the design phase also helps mitigate risk and avoid retrofitting later. Shift left encourages hardware and software teams to collaborate on optimized solutions that balance performance, power and functionality.
When it comes to AI software itself, small language models (SLMs) are compact, lightweight models that consume a fraction of the energy compared to more sophisticated models. Examples like Llama 3 8B, phi-3, Mistral and Gemma are optimized for simpler uses like conversation, summarization and translation. With fewer layers and parameters, they need less data for training, which conserves computing power.
SLMs represent a potential paradigm shift in the AI landscape. By lowering energy requirements they could open up AI for a much wider range of users. While LLMs remain powerful tools for a broad range of applications, SLMs offer efficient, specialized solutions, particularly valuable in resource-constrained environments and specific industry applications.
SLMs and LLMs differ primarily in size, complexity and their respective applications. SLMs are designed to be more compact and efficient, typically containing significantly fewer parameters than LLMs. For example, while LLMs like GPT-4 can have around 1.76 trillion parameters, SLMs might have only a few billion, such as Llama 3 8B, which contains 8 billion parameters. This difference in size makes SLMs more suitable for specific tasks and environments, particularly where computational resources are limited.
What’s Next?
Beyond incremental optimizations to existing architectures, there are more revolutionary approaches on the horizon. Entirely new paradigms have the potential to vastly improve energy efficiency and performance.
Neuromorphic computing mimics how the human brain processes information, using artificial neurons and synapses that consume power only when computing “spikes” occur. This design can boost AI performance even as it lowers energy consumption by a large order of magnitude. Quantum computing leverages the quirks of quantum physics to tackle complex problems. Quantum computing currently requires substantial energy, particularly for maintaining the necessary low temperatures for qubit stability, which often involves cooling systems operating near absolute zero. In the future, as quantum computers scale and overcome current technical challenges, they are expected to handle complex calculations more efficiently than classical supercomputers, particularly in fields like cryptography, optimization and materials science.
Both of these technologies are still in development, but they could become practical within the next three to five years, potentially solving many of the current energy issues in AI. This will require significant ecosystem development and planning. The industry needs to be more proactive and ready this time around than we were when LLMs came online. By embracing energy-efficient solutions and preparing for revolutionary computing paradigms, we can make sure the next wave of major technology disruption doesn’t catch us off guard.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
