Interval

Uniform low-precision conversion reduces weight storage but need not minimize CPU inference latency. We propose a hardware-aware precision-selection procedure that profiles complete neural networks with layer wise FP32 or dynamic INT8 operators and chooses the fastest mask satisfying a validation-accuracy constraint. The method exhaustively evaluates eight precision masks for a three-linear-layer classifier and requires no quantization-aware retraining. Using the 1,797-sample handwritten-digits dataset, three hidden widths (256, 512, 1,024), three initialization seeds, and batch sizes 1 and 16 yields 144 measured configurations. The smallest single-request case retains FP32; intermediate matrix sizes favor selective INT8 on the dominant hidden layer; the widest batch-16 model favors uniform INT8. An independent timing pass with frozen masks and held-out inputs confirms mean speedups of 1.87× at width 512/batch 16 and 4.12× at width 1,024/batch 16, relative to FP32. Selected masks match the paired FP32 held-out test accuracy in all configurations. The results show that small-scale, backend-aware precision profiling can improve CPU inference without optimizing quantization kernels or retraining the deployed model.