That benchmark appears to compare full precision fp32 inference on the nano with uint8 inference on the coral, that floor wiping comes with a lot of caveats