FPGA vs GPU Reality Check
Recent industry pieces and threads are sharpening the FPGA vs GPU trade for aerospace: FPGAs still win for deterministic, sensor‑near processing and custom I/O, while GPUs offer faster iteration and higher throughput for evolving ML models. The conversation now centres on mission partitioning — use FPGAs for tight timing and custom interfaces, GPUs for perception and model‑heavy tasks — rather than an absolute ‘winner’. (x.com, x.com)
A graphics processing unit is like a warehouse full of identical workers doing the same kind of math on huge piles of data at once, which is why NVIDIA sells Jetson modules for edge systems that run perception, language, and multi-camera artificial intelligence models on-device. (nvidia.com) A field-programmable gate array is closer to building the factory itself, because engineers wire the data path so bits move through fixed lanes instead of waiting in line for a general-purpose processor. AMD says its Versal adaptive system-on-chips combine programmable logic, processors, and artificial intelligence engines in one device for heterogeneous acceleration. (docs.amd.com) That difference shows up first in timing. In radar, electronic warfare, and sensor ingest, aerospace teams often care less about peak math and more about whether every packet arrives on the exact cycle they planned, and Curtiss-Wright describes field-programmable gate arrays as the usual low-latency choice for time-critical radar and sensor data. (defense-solutions.curtisswright.com) It also shows up at the connector. A field-programmable gate array can be shaped around custom camera links, radio front ends, SpaceWire-style interfaces, and odd legacy buses in a way a graphics processing unit usually cannot without extra bridge chips or software layers. (arxiv.org) The graphics processing unit wins a different contest. NVIDIA’s Jetson AGX Orin technical brief says the module can deliver up to 275 trillion operations per second for artificial intelligence workloads, which is the kind of throughput teams want when a perception stack keeps changing every few months. (nvidia.com) Software is the second reason graphics processing units keep spreading. NVIDIA’s Jetson Platform Services pitches modular services for building and deploying edge artificial intelligence applications, and that matters because retraining, swapping, and benchmarking models is much faster when the toolchain looks like mainstream machine learning instead of hardware design. (docs.nvidia.com) That is why the argument has shifted from “which chip wins” to “which job goes where.” The United States Small Business Innovation Research program now describes heterogeneous computing systems as mixes of field-programmable gate arrays, graphics processing units, tensor processors, and other accelerators that offload specialized tasks such as machine learning or image processing. (sbir.gov) In practice, the split is becoming pretty clean. Put the field-programmable gate array next to the sensor for framing, synchronization, filtering, and custom input-output, then hand cleaned-up data to the graphics processing unit for object detection, segmentation, tracking, or other model-heavy perception work. (arxiv.org) Space programs are moving the same way. NASA’s High-Performance Spaceflight Computing project says future missions need onboard computing for autonomy, image and signal processing, data-flow management, and object detection, which is exactly the mix that pushes designers toward more than one kind of accelerator. (nasa.gov) Even the hardware roadmaps point in that direction. AMD’s recent space presentation talks about low-latency high-bandwidth streams and on-orbit reconfiguration for adaptive devices, while NVIDIA’s Jetson line keeps pushing larger on-device artificial intelligence pipelines, so aerospace buyers are increasingly matching each processor to the part of the mission it handles best. (indico.esa.int, nvidia.com)
