Klepsydra AI.
Performance Benchmark

August 4, 2020

Overview of the performance benchmark

This report present the results of the Klepsydra AI performance benchmark. The experimental setup consists of the following specifications:

Net: AlexNet
Baseline solutions: OpenCV 3.4.10, TensorFlowLite (TFL).
Processors: Odroid XU4, Raspberry Pi 4
Operating system: Ubuntu 18.04
Frames per second: 0.1, 0.2, 0.25, 0.5, 1, 2, 4

Benchmark criteria:

Power consumption: CPU usage vs processed image rate
Memory usage: RAM usage vs frames per second
Throughout: Processed image rate vs frames per second
Latency: Image processing time vs frames per second.

Analysis of Klepsydra AI Performance Benchmark

The full list of results can be found in the figures below. Here is a summarised list of the results.

Raspberry Pi4

	FPS	Klepsydra	OpenCV	TFL
Power Consumption	0.1	46%	86%	81%
RAM Usage	0.1	1.5Gb	1.6Gb	1.6Gb
Throughput	0.1	0.1	0.08	0.1
Latency (ms)	0.1	630	5000	3200

Table 1.Benchmark at 0.1 Frames per second

	FPS	Klepsydra	OpenCV	TFL
Power Consumption	2	93%	98%	98%
RAM Usage	2	1.3Gb	1.6Gb	1.6Gb
Throughput	2	1.9	0.08	0.2
Latency (ms)	2	950	N/A	N/A

Table 2. Benchmark at 2 Frames per second

Odroid XU4

	FPS	Klepsydra	OpenCV	TFL
Power Consumption	0.1	22%	36%	31%
RAM Usage	0.1	1.1Gb	1.2Gb	1.3Gb
Throughput	0.1	0.1	0.1	0.1
Latency (ms)	0.1	2000	2800	1200

Table 3. Benchmark at 0.1 Frames per second

	FPS	Klepsydra	OpenCV	TFL
Power Consumption	2	53%	70%	57%
RAM Usage	2	1.0Gb	1.2Gb	1.3Gb
Throughput	2	1.55	0.16	0.9
Latency (ms)	2	2300	N/A	N/A

Table 4. Benchmark at 2 Frames per second

Conclusions

Klepsydra Embedded AI outperforms OpenCV and TensorFlow, the market leaders DNN engine for embedded devices. This results are for CPU-only AI inference.
This technology is patent pending. A new version for GPU is currently under development.

Our technology

Written in C++11, our core software is general purpose in nature. It can be installed in a broad spectrum of on-board computers and operating systems. Klepsydra currently integrates with the main software solutions for automotive industry including CANBus, GPU, Linux, RTOS and MATLAB .