Throughput differs by a few percentage points across same-model GPU instances in the cloud because different thermal environments cause thermal throttling to trigger at different times and depths.