The challenge
Datacenter servers are often replaced as a single unit on a common fleet-wide schedule. Their components improve at different rates: the research finds annual energy-efficiency gains of about 38% for GPUs, compared with 18% for CPUs and 12% for memory. A schedule driven by accelerators retires slower-changing components too early, while keeping an entire server longer delays valuable efficiency gains.
Replacement also carries financial and environmental costs. Manufacturing new hardware produces embodied carbon; operating older hardware consumes electricity and produces emissions that depend on the local grid. The right balance varies by component, location, and the operator’s objectives.
The work
We optimize hardware replacement over a 20-year planning horizon, assigning CPUs, GPUs, memory, and storage their own refresh schedules. The work is a collaboration with the University of Pennsylvania and Google.
The optimization framework combines equipment costs, energy efficiency, embodied carbon, electricity prices, and grid carbon intensity. It compares policies that minimize dollars, carbon, or a joint objective that assigns a price to emissions. Public hardware and energy data support the analysis, with sensitivity studies testing slower technology improvement, cleaner grids, and uncertainty in model inputs.
Research findings
- Separate component schedules reduce waste. Refreshing components independently cuts costs by 11% at the same carbon footprint, or carbon emissions by 13% at the same cost, compared with replacing whole servers on a GPU-driven schedule.
- Location changes the best policy. Combining component-specific schedules, local grid conditions, and carbon pricing reduces total cost by 7.1–24.4% across 17 sites. This comparison uses a carbon price of $320 per tonne and a baseline based on GPU advances and fleet-average conditions.
- Hardware lifetime creates a carbon tradeoff. Refresh schedules account for both the carbon from manufacturing new components and the energy required to keep older hardware running. Component efficiency and the local electricity supply determine the balance.
Why it matters
CPUs, GPUs, memory, and storage improve at different rates. Treating each component on its own schedule makes hardware upgrades more cost-effective and reduces their carbon footprint.
The framework gives researchers and operators a way to plan upgrades around their technology forecasts, workloads, and local energy conditions. It also makes the case for modular hardware that lets components be upgraded independently.
