Fine-grained per-vector scaling for neural network quantization

A system, comprising:

• a processor configured to:
• receive a scale factor for each vector of quantized parameters within a single dimension of a multi-dimensional quantized parameter input tensor to provide a plurality of scale factors;
• compute a vector dot-product value for each vector of quantized parameters;
• modify, for each vector of quantized parameters, each vector dot-product value according to a corresponding scale factor to produce a scaled vector dot-product value;
• and
• accumulate the scaled vector dot-product values to produce each element of a multi-dimensional output tensor.

A non-transitory computer-readable media storing computer instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of:

• receiving a scale factor for each vector of quantized parameters within a single dimension of a multi-dimensional quantized parameter input tensor to provide a plurality of scale factors;
• computing a vector dot-product value for each vector of quantized parameters;
• modifying, for each vector of quantized parameters, each vector dot-product value according to a corresponding scale factor to produce a scaled vector dot-product value;
• and
• accumulating the scaled vector dot-product values to produce each element of a multi-dimensional output tensor.