Fraud control: rules for node owners
A renter pays for the machine as it is listed in the marketplace: its GPUs, CPU and the performance its benchmarks showed. Anything that takes part of that away from the renter without them knowing is fraud. We check every node for it automatically, and a node caught twice is locked for good.
What is not allowed
Don't | Why it is fraud |
|---|---|
Run a miner or any other workload outside idle jobs | It keeps running during a rental and takes GPU and CPU the renter paid for |
Run services of your own on the node, such as a blockchain node, a database or a web server | They hold memory the renter paid for; a VM rental, whose memory is reserved in full when it starts, fails to start or is stopped |
Lower the power limit of a GPU during a rental | The renter gets a slower card than the one listed |
Lock or lower GPU core or memory clocks during a rental | Same: less performance than listed |
Run overclocking or tuning tools that change GPU settings on their own | They can change the cards in the middle of a rental, which counts as the same violation |
What is allowed
- Idle jobs. Mining or any other workload set up as an idle job in the dashboard. Idle jobs are stopped automatically before a rental starts and come back after it ends.
- HiveOS miners. A miner started from a HiveOS flight sheet is stopped and started by the platform the same way.
- GPU profiles. Power limit, clocks and fan speed set through a GPU profile in the dashboard. The platform applies the profile only while the node is idle, so a rental always gets the settings it started with.
- Changing hardware. It is not fraud, but the node loses its Verified status, gets verified and benchmarked again, and returns to the marketplace with the new results.
How we detect it
Workloads outside idle jobs
About once an hour, at a random time, an idle node gets a short probe:
- Everything the platform manages is stopped, exactly as before a rental: idle jobs and HiveOS miners.
- After a few seconds, we look at what is still running.
- Any process that still uses 2 GB or more of GPU memory, or CPU load of 50% or more, or any process other than the agent and Docker that holds 2 GB or more of memory, is a violation.
- Idle jobs are started again right away.
Your idle jobs pause for a few seconds during each probe. This is expected.
One detection may be a coincidence, so it is not acted on. Two detections in a row lock the node.
This check covers NVIDIA, AMD and Intel GPUs, and the CPU on every node.
GPU settings changed during a rental
While a compute or render rental runs on a node, the cards are checked every minute against the settings the platform applied before the rental.
What counts as a change:
- Power limit: any change counts.
- Clocks: a card that is at 90% load or more, is not held back by heat or its own power limits, and still runs below 80% of its allowed clock for 3 minutes in a row counts as clock-locked.
When a change is found:
- The renter sees it in the session log.
- The cards are set back to the applied settings. The rental is not stopped.
- The node gets a strike.
One change counts as one strike, however long it lasts. Two strikes take the node out of the marketplace at once (no new rentals), and the node is locked as soon as the current rental ends.
Slowdowns the card causes itself, from overheating or its own power limits, don't count. Keep the cooling in good shape so the cards run at full speed under load.
What happens when a node is locked
- The node is removed from the marketplace and disconnected from the platform. The agent stops reconnecting and writes the reason to its log.
- The lock stays on the machine: reinstalling the system or the agent, or enrolling the same machine again, does not lift it.
- The node's earnings for the current period are held for review. They can be forfeited and refunded to the renters who were affected.
If you believe your node was locked by mistake, contact support with the node number.
Avoid false detections
Mine only through idle jobs or HiveOS flight sheets, never with a miner you started yourself.
Remove miners and other GPU or CPU or memory-heavy workloads that start on their own at boot, Docker containers with a restart policy included.
Turn off overclocking and tuning tools, and set GPU settings only through a GPU profile in the dashboard.
Don't run nvidia-smi -pl, -lgc or -lmc by hand, or scripts that do.
Don't use the node for anything else: no desktop sessions, games or your own jobs on the GPUs.
Keep the cooling clean and working, so the cards don't throttle under load.