Swapping GPUs for FPGAs Could Be a Major Efficiency Boost for AI-Enhanced Telecommunications

Researchers from TU Dresden's Laboratory of Measurement and Sensor System Technique have come up with a way to boost the performance of optical communication systems while simultaneously dropping their power requirements — by replacing GPU-based deep neural network acceleration with a field-programmable gate array (FPGA).

"Mode division multiplexing (MDM) using multi-mode fibers (MMFs) is key to meeting the demand for higher data rates and advancing internet technologies. However, optical transmission within MMFs presents challenges, particularly due to mode cross-talk, which complicates the use of MMFs to increase system capacity," the team explains. "With the success of deep neural networks (DNNs), AI [Artificial Intelligence]-driven mode decomposition (MD) has emerged as a leading solution for MMFs. However, almost all implementations rely on Graphics Processing Units (GPUs), which have high computational and system integration demands. Additionally, achieving the critical latency for real-time data transfer in closed-loop systems remains a challenge."

The team's solution is simple enough: replacing the GPU in the system, a technology that was originally designed to accelerate graphics rendering and has since become the go-to device for all kinds of highly-parallel workloads including machine learning and artificial intelligence, with an FPGA running a gateware designed to accelerate a custom-trained convolutional neural network (CNN). This network, the researchers explain, was trained on synthetic data for one task alone: predicting mode weights from intensity images.

The resulting network is run on the FPGA itself, a low-cost AMD Zynq-7020, using fixed- rather than floating-point arithmetic — and in testing proved capable of identifying up to 30 modes with total accuracy and of performing decomposition for up to six modes with a minimal correlation of around 90 per cent at rates over 100 frames per second. More importantly it did so considerably more efficiently than using a GPU to accelerate the same task, drawing just 2.4W at full load compared to the order-of-magnitude higher power draw of the GPU acceleration approach.

The team's work has been published in the journal Light: Advanced Manufacturing under open-access terms.