Work

Provisionamento de inferência de DNNs com consciência de acurácia para sistemas de borda multi-FPGA heterogêneos

Year: 2026
Type: dissertation
Abstract: The growing adoption of IoT devices at the network edge has established edge computing as an essential paradigm for meeting latency constraints in deep neural network (DNN) inference applications. Heterogeneous multi-FPGA platforms offer flexibility to accelerate these workloads, yet efficiently provisioning inference accelerators on such systems remains challenging due to the trade-off between accuracy, parallelism, and response time. High-accuracy accelerator variants consume more FPGA resources, limiting parallelism, while compact approximate variants increase throughput at the cost of inference quality. Varying inference demand intensifies this tension by creating contention that static provisioning approaches cannot accommodate. This work develops a provisioning framework that composes static accuracy filtering with dynamic contention-aware pool selection in a three-stage pipeline for heterogeneous multi-FPGA edge systems. The first stage applies a static accuracy filter that establishes a quality floor for the design space. The second stage employs an adaptive contention-aware window that dynamically adjusts the eligible design pool based on workload intensity. The third stage optimizes design-to-FPGA allocation and task-to-instance distribution through configurable provisioning strategies. The framework is evaluated under varying inference demand on clusters of 4–8 FPGAs comprising three board types. In simulation, a balanced configuration cuts response time 69% below the highest-accuracy baseline while staying 19 percentage points more accurate than the highest-throughput baseline.
Language: English
Open Access status: green
OpenAlex ID: W7212328653
Cites: 47
Cited by:
Related to: 10