Cloud Accelerator Proximity Map

Link

An interactive web-based tool to visualize the proximity of major cloud provider GPU and accelerator data centers to your own locations. This map helps you find the closest cloud regions with specific accelerator hardware, making it easier to plan for low-latency deployments.

Features

• Interactive World Map: Visualize global data center locations from Google Cloud (GCP), Amazon Web Services (AWS), and Microsoft Azure.
• Hardware-first Filtering: Search GPU models, architectures, vendors, or provider families; select a whole model or individual families with checkboxes. Visible, removable selection chips show the exact hardware and variants. Provider changes and searches never silently erase selections.
• Custom User Locations: Add your own data center or user locations by pasting a simple JSON object to find the nearest cloud regions.
• Proximity Calculation: Adding locations immediately shows the 5 closest matching cloud regions. Switch with the location selector or a green marker. Distances are approximate straight-line kilometers, not measured network latency.
• Shareable State: The current view—including selected provider, accelerator types, and your custom locations—is encoded in the URL. Simply copy the URL to share your exact configuration with others.

Add Your Locations

1. Open the Your Locations tab.
2. Paste a JSON object keyed by location name. Coordinates must be numbers: lat between -90 and 90, and lng (or lon) between -180 and 180.
3. Click Update Map. Green markers appear and proximity results open for the first location (or your previously selected location if it still exists).
4. Use Find regions near or click a green marker to select another location. Compare All Locations gives the nearest matching region per provider.

Only applied, validated locations are saved in the URL; invalid edits leave your previous map intact. Shared URLs contain your location names and coordinates, so share them only with intended recipients.

Example:

{
  "Frankfurt, DE": {
    "lat": 50.1109,
    "lng": 8.6821
  },
  "London, UK": {
    "lat": 51.5074,
    "lng": -0.1278
  },
  "San Francisco, US": {
    "lat": 37.7749,
    "lng": -122.4194
  },
  "Singapore": {
    "lat": 1.3521,
    "lng": 103.8198
  }
}

Basemap

The map uses OpenStreetMap standard tiles with visible attribution, normal browser caching, and an origin-only cross-origin referrer. CARTO's formerly keyless endpoint now returns "API key required" tiles, so it is no longer the default. Use is limited to normal interactive viewing under the OSM tile usage policy: no prefetch, bulk downloads, or offline tiles. The public service has no SLA; a high-traffic deployment should use a suitably provisioned tile provider.

Proximity Regression Tests

Run node --test tests/test_map.js with Node.js 22 or newer. These dependency-free tests run the page's actual script and event handlers with DOM/Leaflet doubles, covering coordinate validation, lon normalization, automatic selection, marker/selector switching, sorted distances, category reset, saved-link restoration, escaped location names, and visible provider-load failures. They do not replace visual browser checks or test live tile delivery.

Hardware Classification and Filtering

accelerator_catalog.json is the explicit family → hardware-model catalog, with provider specification links and a verification date. It distinguishes GB300, GB200, B300 and B200, as well as RTX PRO 4500/6000, A10/A10G, AMD GPUs, and non-GPU ASIC/FPGA/video accelerators. For example, GCP A4X Max maps to GB300, A4X to GB200; AWS P6e-GB300 maps to GB300, while P6-B300 maps to B300. Azure NDasr A100 v4 uses 40 GB GPUs, distinct from NDamsr A100 v4's 80 GB GPUs.

Search narrows the choices, not the map. A model checkbox selects all listed families in the current provider/search scope; expanding a model lets you choose specific families. Multiple choices use OR matching. Selected hardware remains visible outside the filter tab and can be removed individually. Switching provider preserves choices and clearly identifies out-of-scope families; it never falls back silently to all hardware. Existing gpus= shared URLs remain compatible, including removed families, which show a warning and match no regions until cleared.

Provider badges count distinct mapped regions, not machine-family counts or live capacity. Provider-confirmed AWS P6e GB200/GB300 and Azure ND GB200/GB300 offerings remain selectable even when the regional dataset omits them. They show a confirmed provider badge, specification links, and a separate region-coverage notice. The map only places verified locations; it never infers a location from another GPU family. Rankings and comparison exports flag incomplete coverage rather than treating missing region data as unavailable hardware. A source may document hardware even when its region-availability matrix does not list it. Unknown families appear as Unclassified; CI checks classification coverage again after the live refresh so new families require a reviewed catalog entry.

AWS confirms P6e-GB300 on the P6 product page and in its June 18, 2026 PCS announcement. The Capacity Blocks availability table separately lists P6e-GB200 in the Dallas Local Zone (us-east-1-dfw-2a); its location cannot be assumed for GB300.

Data Sources

1. GCP: GPU regions and zones.
2. AWS: Amazon EC2 instance types by Region.
3. Azure: Product Availability Per Region.

Note

Coordinates represent the approximate center of the region or its reference city as listed in the official documentation. Data was last updated October 2026. AWS GovCloud is intentionally outside the map's public/commercial cloud scope.

Refreshing Provider Data

The updater fetches all three official provider sources, validates the complete result, and only then replaces the generated JSON files. It backs up the prior generation and restores it if publishing any provider file fails. Region coordinates and canonical labels live in region_metadata.json, so a newly launched provider region stops the refresh with a clear metadata error instead of producing a partial map.

python -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt
python -m unittest discover -s tests
python scrape_data.py

Azure records retain the map's display slug in region and expose Azure's canonical programmatic name in region_id (for example, east-us-2 and eastus2).

Microsoft documents the ND GB200 v6 and ND GB300 v6 families, but the map does not assign them to regions until Microsoft's product-availability source reports a dependable GA region mapping.

Pull requests and a monthly scheduled GitHub Actions run execute the same tests and refresh. The refreshed aws.json, azure.json, and gcp.json files are attached to the workflow run as an artifact for review, and the job fails if those generated files differ from the committed snapshot.