GPU data health.

How the numbers are protected

• Only fetched numbers. An adapter throws on anything it does not recognise, and nothing is typed in by hand.
• A broken read is refused. A fetch whose machine count falls below half the last one, or whose median rate per GPU-hour moves more than 2.5 times either way, is rejected and the previous reading kept, so a half-broken parser cannot publish nonsense.
• The board is not the provider's to describe. Memory, bandwidth, architecture and launch year come from a registry a person writes. An adapter meeting a board with no entry throws, so a new GPU arrives as a visible failure rather than as a page printing a chip with no specification.
• One number, whatever the machine. Every rate is divided down to one GPU for one hour, which is the only way a single card and a rack compare. The market median today is $1.84/GPU-hr.
• History on the published number. A currency moving is not a price change.

workflow runs on GitHub ↗ · what moved this week → · every provider