A method and device for on-chip bus arbitration are disclosed. The method includes: dividing devices into a first level, a second level and a third level from high to low; and in each arbitration period, calculating remaining processing time of each real-time transaction, and upgrading a device making a request required to be processed immediately to the first level in the current arbitration period; monitoring bandwidth usage amount of devices of the first level and the second level respectively, and downgrading a device whose bandwidth usage amount exceeds a preset bandwidth threshold value to the third level in the current arbitration period; and in devices making requests for a bus use right, if a device of the highest level is the device of the first level, authorizing the device of the first level; and if it is not the device of the first level, authorizing a device making continuous requests.