Hardware Costs Outweigh Gains for Learned Gating in AMC
New FPGA benchmarks show that adding learned feature gating to modulation classifiers increases logic and latency without improving accuracy.
Adding learned feature gating to automatic modulation classifiers (AMC) on FPGA hardware increases resource consumption and latency while actually reducing classification accuracy. In a bit-accurate evaluation using an Intel Cyclone V FPGA, the addition of a 32-element input-dependent gate resulted in a mean accuracy drop of 0.784 percentage points under post-training quantization and 0.616 percentage points under quantization-aware training.
The hardware overhead is concrete. The gating mechanism requires an additional 1,318 adaptive logic modules, 1,557 registers, and 4 DSP blocks. It also adds 3,140 processing cycles to the classification pipeline.
This creates a significant efficiency problem for edge-based signal intelligence. For systems relying on compact fixed-point classifiers—specifically those using sparse DFT-energy and statistical features—the logic cost of gating does not translate into a performance win. The trade-off is negative: more silicon area and higher latency for lower precision.
Companies building high-throughput radio frequency (RF) front-ends and software-defined radios must weigh these findings against software-only simulations. Software models often suggest that feature reweighting improves performance, but this hardware validation shows that those gains can vanish or reverse when ported to fixed-point FPGA implementations.
The inconsistency of quantization-aware training further complicates the design process. The effect of this training method changed direction between different training seeds, suggesting that hardware-level performance in gated models is unstable and highly sensitive to initialization.
Engineers designing for the edge should prioritize leaner, ungated architectures for modulation classification. The data indicates that for this specific feature representation, the simplest path is the most accurate and the fastest.