Researchers say software optimizations can cut AI data center energy use by 30 percent
New findings suggest that adjusting algorithms and precision formats offers a viable path to reducing power consumption without new hardware.
Key points
- University of Michigan researchers found that software changes can significantly lower energy use in AI workloads.
- Running inference in FP8 format consumed one-third less energy than bfloat16 on specific problem-solving tasks.
- A custom training optimizer reduced energy consumption by up to 30 percent while maintaining throughput levels.
Researchers from the University of Michigan indicate that software modifications could substantially reduce the energy consumption of artificial intelligence infrastructure. Jae-Won Chung, a PhD candidate and researcher with the ML.Energy initiative, stated that algorithms and systems software can meaningfully help with power management, according to Tom's Hardware.
The study highlights that current industry efforts focus heavily on hardware efficiency, such as building better GPUs and securing grid connections. However, Chung argues that the hardware layer is difficult to change quickly. In contrast, systems software, algorithms, and applications are easier to modify, allowing efficiency gains to compound across the computing stack without replacing physical components.
Context and Background
Data center energy demand is rising sharply, with projections indicating that AI workloads will consume approximately 945 terawatt-hours of electricity by 2030. The International Energy Agency notes this amount equals the current total electricity usage of Japan. Despite this surge, the Uptime Institute’s 2025 survey shows that average power usage effectiveness has remained largely unchanged for six consecutive years.
Power usage effectiveness measures cooling and power system waste but does not account for whether server software performs useful work. Servers account for roughly 60 percent of electricity demand in modern data centers. Cooling systems consume between 7 percent and 30 percent of power, depending on facility efficiency. Therefore, optimizing the software layer represents a significant opportunity for conservation.
Who Is Affected
The findings directly impact hyperscale data centers and enterprises managing large-scale AI training and inference workloads. Major GPU manufacturers are already exploring similar software-side efficiencies, indicating a broader industry shift toward algorithmic optimization. The research demonstrates that lowering numerical precision during inference can yield substantial energy savings without degrading performance on problem-solving tasks.
Specific tests on the Alibaba Qwen 3 235B A22B Thinking model showed that using the FP8 format consumed one-third less energy compared to bfloat16 versions. Additionally, Chung’s Perseus training optimizer identified underutilized parts of large-model training jobs and slowed them down. This synchronization allowed tasks to finish alongside busier components, cutting training energy by up to 30 percent without altering hardware or reducing throughput.
What should you do?
Organizations should evaluate their current AI workload configurations to identify opportunities for precision reduction. Administrators can review training pipelines to detect idle or underutilized processes that could be throttled. Monitoring tools should be adjusted to track software-level efficiency metrics alongside traditional power usage effectiveness data. Teams should also stay informed about updates from GPU vendors regarding software optimization features.
What to do and how to stay safe: AI
- Audit current AI inference workloads to determine if lower-precision formats like FP8 can be applied without impacting accuracy requirements.
- Review training job architectures to identify segments with low computational load that can be slowed to synchronize with busier tasks.
- Implement monitoring solutions that track software-level efficiency gains in addition to standard hardware power metrics.
- Consult with vendor documentation to understand available algorithmic optimization features in current GPU firmware and drivers.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How much energy can software optimization save in AI training?
The Perseus training optimizer reduced energy consumption by up to 30 percent without changing hardware or reducing throughput.
What is the projected electricity demand for AI by 2030?
AI is expected to use approximately 945 terawatt-hours of electricity by 2030, according to industry projections.
Does lowering precision affect model performance?
Tests on the Alibaba Qwen 3 model showed that FP8 format consumed less energy on problem-solving tasks without stated performance degradation.
