However, low-precision computing is proving that many tasks don’t need 32-bit floats. Speech recognition models, for instance, now run efficiently in INT8 with minimal loss in accuracy.