NNS: Neural Network Dimensioned by Shannon Theorem

Abstract: We introduce NNS (Neural Network dimensioned by Shannon Theorem), a method producing two independent neural outputs from a single GPU dispatch and a single weight buffer load. The second channel is derived from the bitwise transpose of the packed weight block, computed on-the-fly in shared memory at zero storage cost. Validated on commodity mobile GPU via standard OpenGL ES 3.1 without root access or proprietary SDK, NNS achieves 1.91x-2.16x effective throughput improvement across four quantization modes, with overhead ranging from -7.3% (dual faster than single) to +4.8%. Memory overhead is 31%, confined to a single reutilizable output buffer.