“The trend in neural network arithmetic is toward lower precision, with 4-bit precision becoming a primary format, similar to NVIDIA's approach.”