Sept. 17, 2020 — The Barcelona Supercomputing Center (BSC) and EmbeDL AB, a spin-off of the BSC coordinated project LEGaTO, come together to enable low-energy optimizations for the embedded Deep Learning (DL) applications in the context of the LV-EmbeDL project, a Bilateral Technology Transfer Experiment (TTX) granted by Tetramax.
EmbeDL AB is an AI research spin-off that can help organizations adopt DL methods into their applications, like autonomous cars, drones, and even smart homes, significantly faster while using less energy and memory. At the heart of this technology is the DL optimization engine, EmbeDL, developed in the LEGaTO use case in machine learning.
LV-EmbeDL aims to equip EmbeDL with BSC’s FPGA Undervolting technique developed in the LEGaTO project. Undervolting is an effective technique explored and developed by researcher Behzad Salami, to maximise energy efficiency through supply voltage underscaling below the default level of the underlying hardware.
“I am very pleased that we can now take a practical step to put the FPGA Undervolting technique developed in the LEGaTO project into practice, thanks to Tetramax. We have already shown the significant efficiency of this technique in research papers (for example, in MICRO2018 and DSN2020). However, within the LV-EmbeDL project and in collaboration with EmbeDL AB, I am looking forward to taking a big step further and developing it in the industrial environment”, said Behzad Salami (BSC), Principal Investigator (PI) of LV-EmbeDL.
EmbeDL optimizes DL models to run faster and more energy-efficiently as well as to reduce development and hardware costs. EmbeDL achieved an average of 5X more energy efficiency on the same hardware by exploiting techniques like quantization and pruning. In LV-EmbeDL, we aim to reach close to 50X energy-efficiency improvement on the same hardware thanks to the Undervolting technique, which would be a truly extraordinary achievement.
LV-EmbeDL will be demonstrated in a drone perception system in an operational environment running EmbeDL for pixel-wise semantic segmentation of an RGB camera stream. Industry partner’s customers will utilise it in domains like autonomous cars, telecom, and IoT.
LV-EmbeDL received the I4MS-SAE Label, recognising the project’s “excellent implementation, high potential for further deployment and innovative aspect.” ICT Innovation for Manufacturing SMEs (I4MS) and Smart Anything Everywhere (SAE) are initiatives promoted by the European Commission to foster the digital innovation of SMEs in Europe and to boost their competitiveness.