Table 6 Baseline results for the acoustic scene classification task

From: The neurobench framework for benchmarking neuromorphic computing algorithms and systems

Baseline

Accuracy

Execution Time (ms)

Idle Power (mW)

Active Power (mW)

Dynamic Power (mW)

Dynamic Energy (mJ/inf)

CPU

pre-process

79.64%

43

79.40

100.72

21.32

0.917

(system-wise)

inference

 

45

79.40

100.15

20.75

0.934

Xylo

pre-process*

79.90%

-

0.00017*

0.015*

0.015*

0.015*

(component-wise)

inference

 

84

0.351

0.692

0.341

0.028

  1. Pre-processing of the Xylo (marked with an asterisk *) measures power of the analog pre-processor in real-time relative to the audio data, whereas other measurements are of digital components processing digital data on-hand. Idle Xylo pre-processing measures silence and active measures test data audio played to the device, and energy is measured over the sample duration of 1 second. CPU power is measured over the full Arduino system, as it does not have on-board power instrumentation, while Xylo power measures power consumed by the Xylo Audio 2 ASIC only. Dynamic power and energy provide proper comparison between the systems, and idle and active measurements are provided for transparency.
Back to article page