Overview of Klepsydra AI

In order to process all the amount of data coming from all the different sensors in a car, Klepsydra SDK is the perfect solution to boost data processing with no data loss. However, for autonomous navigation the use of AI algorithms will be critical. For that reason Klepsydra has developed an AI framework called Klepsydra AI.
Klepsydra AI is a high performance deep neural network engine for edge computers. Customers can deploy existing or new trained models on the edge using Klepsydra AI, in the same manner as with standard edge AI solutions. Klepsydra AI offers three main benefits:

• Process up to 5x more data with AI
• Reduce power consumption
• Compatible with main providers of AI software.

Klepsydra AI can deploy on the edge pre-existing or new models that customers might have developed and trained.

Core benefits of AI for autonomous navigation

Safety and reliability
Real-time edge AI. Klepsydra AI can process data in real time with low latency.
Predictable edge AI. Klepsydra AI is substantially more stable, predictable and deterministic than other edge solutions.

Cost

Less hardware cost. With Klepsydra AI, more data can be processed on the same hardware, with less power and less memory and without any cloud computing support.

Compatibility

Klepsydra AI is compatible with most AI formats and AI software solutions. It also accept input data in a variety of formats including images and time series.
Furthermore, Klepsydra can be deploy to several edge computers including, but not limited to, Odroid XU4, RaspberryPi4, Intel NUC, etc.

Figure 1. Klepsydra AI setup.

Technical specs

Overview

Klepsydra AI is an inference engine for Deep Neural Networks (DNN) aimed at Edge computing applications. Klepsydra AI has the following modules
and APIs:

• Support for ONNX and proprietary reduced file size format (KBAIF).
• API in C++, C and Java
• Self-tuning software to calculate best configuration for low latency, low power consumption and/or high processing data rate.

Core features

Klepsydra has four main variations:

• Standard. Full-fletched version of the software.
• Minimal. 4Mb footprint binary Legacy. Support to older operating systems and compilers.
• Minimal-Legacy. 4Mb footprint binary with support to older operating systems and compilers.

Figure 2. Power consumption benchmark comparison for AlexNet DNN on RaspberryPi4.

Requirements

Klepsydra is platform independent with the following technical requirements:

• Linux, Linux with RT patch, FreeRTOS 10+, RTEMS5+
• Target processor with atomic operation set, including ARM Cortex 32 and 64 bits, x86 64 and 32 bits. Please request Support Matrix for full list.
• C++11 complier for the target computer
• Vectorization and FPU instructions (NEON, AVX)

Figure 3. Processed data volume benchmark comparison for AlexNet DNN on RaspberryPi4.

Compatibility features

• Supported languages: C++, C, NodeJS, Python
• Data format: Float matrices, time series and OpenCV matrix objects (cvMat)
• Model format: ONNX, TensorFlow and Caffe.