Real-Time Fusion Throughput on Spatial FPGA Fabric

Abstract

Greg Passmore

This paper provides a technical assessment of CPUs, GPUs, and FPGAs for real-time sensor fusion workloads involving radar, EO/IR imagery, LiDAR, inertial sensors, and RF data. The analysis covers throughput, latency, determinism, power efficiency, interconnect behavior, and support for large numbers of asynchronous sensor inputs. Empirical studies and published measurements show that CPUs offer flexibility but limited throughput and poor scaling under high-rate multi-sensor ingest. GPUs deliver the highest theoretical parallel throughput but depend on large batch sizes, incur millisecond-scale latency, and rely on PCIe transfers that constrain fusion performance. FPGAs provide deterministic, microsecond-scale processing, spatial parallelism for concurrent sensor pipelines, nanosecond-level timestamping, and significantly higher energy efficiency for structured DSP and streaming workloads. Benchmarks across FIR filtering, vision kernels, and real-time inference demonstrate FPGA speedups ranging from modest to substantial relative to CPUs and GPUs, with consistent gains in throughput per watt. FPGAs also mitigate interconnect bottlenecks by performing in-situ reduction before data reaches GPUs or CPUs. My analysis indicates that heterogeneous architectures, with FPGAs as front-end fusion engines and GPUs and CPUs as downstream analytic and control elements, provide the most effective performance profile for modern sensor fusion systems.