Evaluating deep reinforcement learning against classical 3GPP Mode 2 sensing-based scheduling under extreme vehicle densities.
In simulated intersections with 300 transmitting vehicles per square kilometer, AI-scheduled MAC maintained 98.2% packet delivery ratio, whereas classical Mode 2 dropped to 71.4%.
Dedicated neural accelerator blocks draw only 1.2 watts of additional power on automotive telematics chipsets, easily fitting within thermal budgets.