[Paper Review] Are FPGAs Suitable for Edge Computing?

Saman Biookaghazadeh, Fengbo Ren+1 more|arXiv (Cornell University)|Apr 17, 2018

Advanced Adaptive Filtering Techniques2 references16 citations

TL;DR

This paper evaluates FPGAs for edge computing by comparing throughput, energy efficiency, and adaptability against GPUs using OpenCL-based kernels on an Intel Arria 10 FPGA and NVIDIA Tesla K40m GPU. Results show FPGAs deliver consistent throughput regardless of workload size, outperform GPUs by up to 30.7× in energy efficiency, and maintain stable performance across high-dependency algorithms due to fine-grained spatial and temporal parallelism.

ABSTRACT

The rapid growth of Internet-of-things (IoT) and artificial intelligence applications have called forth a new computing paradigm--edge computing. In this paper, we study the suitability of deploying FPGAs for edge computing from the perspectives of throughput sensitivity to workload size, architectural adaptiveness to algorithm characteristics, and energy efficiency. This goal is accomplished by conducting comparison experiments on an Intel Arria 10 GX1150 FPGA and an Nvidia Tesla K40m GPU. The experiment results imply that the key advantages of adopting FPGAs for edge computing over GPUs are three-fold: 1) FPGAs can provide a consistent throughput invariant to the size of application workload, which is critical to aggregating individual service requests from various IoT sensors; (2) FPGAs offer both spatial and temporal parallelism at a fine granularity and a massive scale, which guarantees a consistently high performance for accelerating both high-concurrency and high-dependency algorithms; and (3) FPGAs feature 3-4 times lower power consumption and up to 30.7 times better energy efficiency, offering better thermal stability and lower energy cost per functionality.

Motivation & Objective

• To evaluate the suitability of FPGAs for edge computing in comparison to GPUs.
• To analyze throughput predictability across varying workload sizes.
• To assess energy efficiency and thermal stability under diverse algorithmic workloads.
• To investigate architectural adaptiveness to concurrency and data dependency.
• To determine if FPGAs can serve as a viable alternative to GPU-based edge servers.

Proposed method

• Benchmarking FPGA and GPU performance using OpenCL kernels on an Intel Arria 10 GX1150 FPGA and NVIDIA Tesla K40m GPU.
• Measuring raw and normalized throughput across varying batch sizes and data dependency levels.
• Recording real-time power consumption using nvidia-smi and Nallatech APIs for energy efficiency calculation.
• Evaluating performance on matrix multiplication and conditional dependency workloads with 8 to 1024 operations.
• Analyzing throughput and energy efficiency across low- and high-dependency algorithmic scenarios.
• Using high-level synthesis (HLS) to enable portable, high-performance FPGA kernels in C-based OpenCL.

Experimental results

Research questions

• RQ1 Can FPGAs maintain consistent throughput across varying application workload sizes, unlike GPUs?
• RQ2 How does FPGA energy efficiency compare to GPUs across different data concurrency and dependency levels?
• RQ3 To what extent can FPGAs adapt to high-dependency algorithms through dynamic hardware reconfiguration?
• RQ4 Does FPGA performance degrade under conditional branching or complex control flow, as seen in GPUs?
• RQ5 What is the energy efficiency advantage of FPGAs in real-world edge workloads with streaming data?

Key findings

• FPGAs deliver consistent throughput invariant to workload size, while GPU throughput drops significantly with increasing workload size.
• FPGAs achieve 2.5 to 30.7 times better energy efficiency than GPUs, with the highest improvement (up to 30.7×) for small batch sizes.
• For high-data-dependency workloads (dependency degree 256), FPGAs achieve up to 11.8 times higher energy efficiency than GPUs.
• GPU performance degrades by up to 37.12 times under conditional branching due to warp divergence, while FPGA performance remains stable.
• FPGAs maintain high throughput across both high-concurrency and high-dependency algorithms due to fine-grained spatial and temporal parallelism.
• FPGAs consume 2.79–3.92 times less power than GPUs, enabling better thermal stability and lower cooling and energy costs in edge deployments.