Hybrid GPU/FPGA Architectures Gain Traction
The trend of marrying GPUs and FPGAs into hybrid architectures for edge AI is accelerating, according to a recent technical presentation. The approach leverages FPGAs for deterministic, low-latency tasks like sensor pre-processing, while offloading high-throughput ML inference to GPUs. This partitioning is becoming a go-to solution for certifiable, real-time AI in aerospace.
The push for hybrid architectures stems from the distinct, complementary strengths of each technology. FPGAs provide deterministic, low-latency processing ideal for the real-time demands of flight control and sensor fusion. Their reconfigurable nature allows for rapid prototyping and in-field updates to adapt to new standards or mission parameters. GPUs, conversely, excel at high-throughput parallel processing, making them highly efficient for compute-intensive AI and machine learning algorithms. This division of labor allows developers to use the optimal processor for each part of the application, from ingesting sensor data with an FPGA to performing complex AI inference on a GPU. Key hardware players are leaning into this trend with System-on-Chip (SoC) solutions. AMD's Versal AI Edge series, for example, combines programmable logic with AI engines and scalar processors for sensor fusion and real-time control. Similarly, Intel's Stratix 10 NX FPGAs integrate AI Tensor Blocks, offering a significant boost in integer performance for AI workloads compared to previous generations. This hybrid approach also offers a clearer path to safety certification. FPGAs have a well-established route for meeting the stringent DO-254 "Design Assurance Guidance for Airborne Electronic Hardware" standard. Certifying complex, multi-core GPUs and their software drivers to the equivalent DO-178C standard for software can be prohibitively expensive and complex. By offloading safety-critical functions to the FPGA, the GPU's certification burden can often be reduced.
