Transceiver bandwidth limits the resolution of ultra-low-power pulse radar systems. Compressed sensing techniques improve resolution but existing efforts require heavy computation. This work proposes a neural-network based compressed sensing radar processor architecture that improves resolution by 6× while remaining computationally efficient. Fabricated in 16nm FinFET CMOS, the processor simultaneously achieves more than 8× throughput and 18× efficiency over the state-of-the-art.