Documentation: Wattlytics -- Analytics for Smarter Energy Decisions
What is Wattlytics?
Wattlytics is an interactive tool that helps data center planners and computational scientists evaluate GPU configurations based on performance per Total Cost of Ownership (TCO). This page provides a detailed breakdown of the tool's purpose, core formulas, and adjustable parameters. Use it as a reference when fine-tuning inputs or interpreting output results.
How It Works
You input your budget, hardware, energy, and operational assumptions. The tool uses benchmark data (e.g., GROMACS or AMBER) to compare GPUs and estimate how much compute performance you get per euro spent over the lifetime of the system.
Input Parameters
Benchmark Setup
- Select Workload: Choose between GROMACS or AMBER benchmarks.
- Benchmark ID: Refers to specific GPU benchmark cases (1–11) based on public performance data.
Initial Setup
- Total Budget (€): The capital available for all GPU setups.
- Same GPU Count: Toggle to force equal GPU quantity across models.
Capital Cost Parameters
- Node Server (€): Cost of server including CPU, memory, storage, networking.
- Node Infrastructure (€): Power distribution and cooling infrastructure cost.
- Node Facility (€): Cost of data center space and construction.
- Software (€): One-time software stack and license costs.
- GPU Cost Mode: Choose how GPU prices are determined.
- Static Prices: Default mode. Uses hardcoded base prices for each GPU.
- Live Delta Prices: Attempts to fetch up-to-date GPU prices from Delta Computer when the "Refresh Live Prices" button is manually triggered. If fetching fails (i.e., "Not Available"), static prices are used automatically to ensure the user never sees a missing value.
Operational Cost Parameters
- Electricity (€/kWh): Cost of electricity.
- PUE: Power Usage Effectiveness (includes cooling overhead).
- Maintenance (€/year): Ongoing node-level maintenance costs.
- System Usage (hrs/year): Expected annual usage in hours.
- Lifetime (years): Projected operational life of the system.
- Node Baseline Power (W): Server power draw excluding GPUs.
- Depreciation (€/year): Annual depreciation (if tracked financially).
- Software Subscription (€/year): Recurring license fees.
- Utilization Inefficiency (€/year): Expected loss due to under-utilization.
- Heat Reuse Revenue (€/kWh): Revenue recovered from heat recovery.
- Heat Reuse Factor (0–1): Portion of heat captured for reuse.
Outputs
- Results Table
- GPU-by-GPU breakdown of performance, cost, and TCO.
- Results are sorted by highest performance per unit cost.
- All output tables are downloadable in CSV format.
- Comparison Message: Summary of best vs worst GPU configuration.
- Scenario Comparison (what-if Simulation): Show side-by-side summaries of two configurations in localStorage and display delta changes in input and output parameters.
- Input Charts for visual insights:
- GPU × Benchmark Performance and Power Heatmaps: Visual representation of performance and power metrics across different GPUs.
- Price Change History Chart: it shows how GPU live prices change over time compared to static prices, helping track trends and cost-performance shifts. Each point reflects: % Difference = ((Live − Static) / Static) × 100.
- Result Charts for visual insights
- Bar Chart: Performance per Total Cost of Ownership (TCO) and the number of GPUs comparison across GPU setups.
- TCO Breakdown: Stacked Bar Chart for each splited TCO cost and Pie Chart for capital and operational costs
- Parameters Sensitivity Analysis: Heatmaps and Tornado Plots
- Output Metrics
- Performance per TCO (ns/day*atoms/€): Higher is better!
- Power per TCO (W/€) : Lower is better! Given a specific budget for a cluster, how much power will it burn? For example, if it exceeds the power limit, adjust the power cap or CPU/GPU clocks until switching to a different GPU is necessary.
- Performance per Watt per TCO (ns/day*atoms/kW/€): Given a specific budget for a cluster, how much work/Joule will it burn?
Tips & Sensitivity Analysis
- Tips: Real-time recommendations appear when inputs are adjusted.
- Sensitivity Analysis: Normalised sensetivity analysis highlights which parameters most affect each GPU’s cost-efficiency.
Need More Help?
Check out the FAQ page for common questions.