arXiv subjects

Frédéric Pétrot

Publications and source records attributed to Frédéric Pétrot.

3 recordsLinked to original sources

arXiv2026

Pipeline Automation Framework for Reusable High-throughput Network Applications on FPGA

In a context of ever-growing worldwide communication traffic, cloud service providers aim at deploying scalable infrastructures to address heterogeneous needs. Part of the network infrastructure, FPGAs are tailored to guarantee low-latency and high-throughput packet processing. However, slowness of the hardware design process impairs FPGA ability to be part of an agile infrastructure under constant evolution, from incident response to long-term transformation. Deploying and maintaining network functionalities across a wide variety of FPGAs raises the need to fine-tune hardware designs for several FPGA targets. To address this issue, we introduce PAF, an open-source architectural parameterization framework based on a pipeline-oriented design methodology. PAF (Pipeline Automation Framework) implementation is based on Chisel, a Scala-embedded Hardware Construction Language (HCL), that we leverage to interface with circuit elaboration. Applied to industrial network packet classification systems, PAF demonstrates efficient parameterization abilities, enabling to reuse and optimize the same pipelined design on several FPGAs. In addition, PAF focuses the pipeline description on the architectural intent, incidentally reducing the number of lines of code to express complex functionalities. Finally, PAF confirms that automation does not imply any loss of tight control on the architecture by achieving on par performance and resource usage with equivalent exhaustively described implementations.

cs.AR↗

arXiv2017

Ternary Neural Networks for Resource-Efficient AI Applications

The computation and storage requirements for Deep Neural Networks (DNNs) are usually high. This issue limits their deployability on ubiquitous computing devices such as smart phones, wearables and autonomous drones. In this paper, we propose ternary neural networks (TNNs) in order to make deep learning more resource-efficient. We train these TNNs using a teacher-student approach based on a novel, layer-wise greedy methodology. Thanks to our two-stage training procedure, the teacher network is still able to use state-of-the-art methods such as dropout and batch normalization to increase accuracy and reduce training time. Using only ternary weights and activations, the student ternary network learns to mimic the behavior of its teacher network without using any multiplication. Unlike its -1,1 binary counterparts, a ternary neural network inherently prunes the smaller weights by setting them to zero during training. This makes them sparser and thus more energy-efficient. We design a purpose-built hardware architecture for TNNs and implement it on FPGA and ASIC. We evaluate TNNs on several benchmark datasets and demonstrate up to 3.1x better energy efficiency with respect to the state of the art while also improving accuracy.

cs.LG↗

arXiv2016

Facing ADAS validation complexity with usage oriented testing

Validating Advanced Driver Assistance Systems (ADAS) is a strategic issue, since such systems are becoming increasingly widespread in the automotive field. ADAS bring extra comfort to drivers, and this has become a selling point. But these functions, while useful, must not affect the general safety of the vehicle which is the manufacturer's responsibility. A significant number of current ADAS are based on vision systems, and applications such as obstacle detection and detection of pedestrians have become essential components of functions such as automatic emergency braking. These systems that preserve and protect road users take on even more importance with the arrival of the new Euro NCAP protocols. Therefore the robustness and reliability of ADAS functions cannot be neglected and car manufacturers need to have tools to ensure that the ADAS functions running on their vehicles operate with the utmost safety. Furthermore, the complexity of these systems in conjunction with the nearly infinite number of parameter combinations related to the usage profile of functions based on image sensors push us to think about testing optimization methods and tool standards to support the design and validation phases of ADAS systems. The resources required for the validation using current methods make them actually less and less adapted to new active safety features, which induce very strong dependability requirements. Today, to test the camera-based ADAS, test vehicles are equipped with these systems and are performing long hours of driving that can last for years. These tests are used to validate the use of the function and to verify its response to the requirements described in the specifications without considering the functional safety standard ISO26262.

cs.CY↗