How Much Energy Does AI Model Training Use—And What Does It Cost?
From Our Knowledge Base
(Costs are per cycle…)
Training an advanced AI model like GPT-4 isn’t just expensive – it’s incredibly energy-intensive. The process requires massive computational power, running for months across thousands of high-performance GPUs. But how much energy does this actually use, and what does it cost?
Let’s break down the energy consumption and cost per stage of a single full training cycle, using an average electricity rate of $0.15 per kilowatt-hour (KWh).
Energy usage in A single training cycle
1. Pre-Processing:
3 Months, $2.25M in Energy Costs
Before training can even begin, terabytes of raw data must be cleaned, filtered, and structured for the model. This phase consumes about 15 million KWh of electricity, leading to an energy bill of $2.25 million—for just one cycle.
2. Post-Preprocessing: 1.5 Months, $675K in Energy Costs
Once the data is refined, additional processing is required, such as feature selection and embedding generation. This step uses 4.5 million KWh, costing $675,000 in energy for a single training cycle.
3. Model Training: 4 Months, $6M in Energy Costs
This is the most power-hungry phase, where thousands of GPUs run nonstop, optimizing billions of model parameters. The total energy use? A staggering 40 million KWh, translating to $6 million in electricity costs per training cycle.
4. Fine-Tuning & Evaluation: 2 Months, $1.5M in Energy Costs
After the core training, the model undergoes fine-tuning using Reinforcement Learning from Human Feedback (RLHF). This step refines its responses and ensures alignment with user needs, consuming 10 million KWh, at a cost of $1.5 million per cycle.
Total Energy Cost of a Single AI Training Cycle: $10.425 Million
From start to finish, training a model like GPT-4 for just one full training cycle consumes 69.5 million KWh of electricity, costing nearly $10.5 million in energy alone. For perspective, that’s enough power to supply thousands of U.S. homes for an entire year.
We want to hear from you.
We know that Augmetrics® is not a universal solution to sustainability problems that we face, but we also know it is a start; one that took over 10 years to develop.
