Loading benchmark data
Every cloud region where AWS, Azure or Google Cloud publishes an on-demand price for a tracked GPU, coloured by what it costs there. The same figures as the benchmark, placed on the ground.
Each outline is a cloud region, drawn as the area its provider names: a built-up area where it names a city, a state or a country where that is all it names. A cloud region is a group of facilities rather than a single datacenter, and no provider publishes their addresses, so nothing here marks a building.
One cloud region, drawn as the area its provider actually names. AWS usually names a city, so it gets a built-up area. Azure often names only a state, so it gets the state. The outline is the region here, and it is what you click.
Price. With one GPU selected it is the price per GPU-hour there. Across all of them it is how that region compares to the world average for the GPUs it actually sells, so a region stocking only A100 is not mistaken for a cheap one.
Buildings. A region is a group of facilities whose addresses no provider publishes, and nothing on this map marks one. Capacity is absent for the same reason: no provider says how many GPUs a region holds, so nothing here estimates it.
Prices are Linux on-demand list prices for the eight-GPU node, normalized to one GPU-hour, and are the same figures behind the market terminal. Sources: AWS, Azure, Google Cloud. Basemap © CARTO, © OpenStreetMap contributors.