This would of course have to be tested but there might be less overhead? I haven’t used FFI but they mention that it is significant. I have written a V8 add-on using NAN and the performance is very good (as in passing relatively large arrays between C++ and Node is in the few microsecond range)