Building a green supercomputer in the cloud
summary
Builds an HPC cluster in the cloud for Vestas that scales to 400,000 cores and handled ~6.5 million IOPS, combining VM-based HPC and serverless for bursts and scale-to-zero. Uses spot/reserved instances (up to ~90% discounts), Carbon Intensity APIs (ElectricityMap), and a carbon-economy SDK to do carbon-aware scheduling. Time-shifting cuts emissions ~12–16%, moving jobs regionally cuts 30–90%, and automated data staging and blob-triggered transfers keep costs and carbon low.
AI-written summary · about
chapters (11)
- 0:00🚀 Introduction: The need for sustainable computing and how it impacts climate change
- 3:00Vestas wind turbines: Engineering challenges and the need for simulation
- 7:03Building a high-performance compute (HPC) cluster for wind turbine simulation
- 11:00Ensuring efficiency: Reducing costs and carbon emissions through cloud optimization
- 14:05The role of programming languages, CPU/GPU selection, and hardware efficiency in green computing
- 18:00Carbon awareness: How to calculate and reduce carbon emissions in cloud computing
- 23:11Using tools like the Green Software Foundation and Carbon Intensity APIs
- 28:44Shifting workloads between regions and utilizing spot instances to reduce carbon footprint
- 32:02Implementing time-shifting to optimize carbon savings in cloud-based workloads
- 37:20Real-world impact: How Vestas saved carbon emissions and costs by optimizing workloads
- 41:10Closing thoughts: How you can apply these practices to make your own cloud infrastructure greener
related talks
58mBuilding a green supercomputer in the cloud
36mCarbon-Aware Computing: Measuring and Reducing the Carbon Intensity of Software
26mAccelerating Science, Decelerating Carbon: Sustainable Computing...
27mGreenOps Sustainable engineering 1
16mA novel Cloud High Performance Computing infrastructure for user
44mCloud computing and carbon emissions
49mBuilding a zero-waste compute infrastructure on
47mCloud Native and Sustainable, Reproducible Scientific Computing
