Konferenzen und Rahmenprogramm
How to Meaningfully Benchmark Edge-Inference Performance in the New Age of Hardware Acceleration
EEMBC created the MLMarkâ„¢ benchmark to standardize inference-performance metrics on so-called "edge devices," devices with a small thermal-design power (TDP) envelope. Compared to historical benchmarks targeted at CPUs, machine-learning benchmarks differ significantly. First, the key element in ML/DL is the graph, not C/C++ source code. Second, the translation of the graph to the hardware depends on the framework, which exposes optimizations for tuning the graph, and the SDK which may contain proprietary strategies for configuring the hardware resources. Third, machine-learning accuracy generally decreases as performance increases due to optimizations. Therefore, a performance score is meaningless without its associated accuracy score. Further complications include parallelism€”such as concurrency and batching€”which have a nuanced difference depending on the application. This paper will explain our methodology for addressing these variables, as well as results and future-proofing.
--- Datum: 25.02.2020 Uhrzeit: 12:00 - 12:30 Uhr Ort: Conference Counter NCC Ost