Modern RF signal intelligence stacks increasingly rely
on ensembles of neural and classical models to stabilize
performance under changing channel conditions, hardware
front-ends, and signal mixes. Majority, weighted, and stacked
voting schemes can suppress idiosyncratic model failures, but
each additional model adds computation, memory traffic, and
host–device synchronization overhead.
In resource-constrained deployments—battery-powered field
nodes, embedded radios, or shared datacenter GPUs with strict
latency service-level agreements (SLAs)—these costs manifest
as a hard cap on the number of signals that can be analyzed
per second. Understanding how latency and energy scale with
ensemble size is therefore critical for deciding whether “just
add another model” is operationally viable.
This paper focuses on a concrete question: given a fixed
pool of RF modulation models, what is the latency/energy cost
of increasing the ensemble size on CPU and GPU, and where
is the “knee” beyond which accuracy gains diminish?