Practical Efficiency of Multiply-Free Ternary Neural Networks: An Empirical Study on Consumer Hardware

This paper presents comprehensive wall-clock benchmarks for training and inference of ternary {-1, 0, +1} state-space models on consumer GPUs (RTX 5070, 12 GB GDDR7). A six-technique optimization stack achieves 2.2x cumulative training throughput improvement at 6M scale. A custom Triton kernel for multiply-free ternary GEMV inference achieves 1.63x speedup and 79% VRAM reduction versus eager-mode Python, though it remains 1.55x slower than FP16 cuBLAS. Optimizer memory reduction from 12 bytes/param (FP32 AdamW) to 2.25 bytes/param (BF16 FiraAdamW) enables 4x model capacity on the same hardware. SSM recurrent state requires 66x less memory than Transformer KV-cache at 50M scale. All results include honest gap analysis identifying specific engineering bottlenecks.