Puce d'accélération de réseaux de neurones

Un Accélérateur d'IA pour accélérateur d'intelligence artificielle (ou NPU, Neural Processing Unit) est une catégorie de microprocesseur ou de systèmes de calculs conçu pour accélérer un réseau de neurones artificiels, accélérer des algorithmes de vision industrielle et d'apprentissage automatique pour la robotique, l'internet des objets et autres taches de calculs-intensifs ou de contrôle de capteurs. Il s'agit souvent de conceptions multicœurs et se concentrant généralement sur l'arithmétique de faible-précision, des nouvelles architectures de flux de données ou de la capacité de calcul en mémoire. Il existe plusieurs termes propres au fournisseur pour certains appareils dans cette page. Historique de puces d'accélération d'IA Les systèmes informatiques ont souvent complété le CPU par des accélérateurs spéciaux pour des tâches spécialisées, notamment des cartes vidéo (GPU) pour les graphiques, mais aussi des cartes son pour l'audio, etc. Au fur

Pyramic array: An FPGA based platform for multi-channel audio acquisition

Juan Azcarreta Ortiz

Microphone arrays techniques present compelling applications for robotic applications. Those techniques can allow robots to listen to their environment and infer clues from it. Such features might enable capabilities such as natural interaction with humans, interpreting spoken commands or the localization of victims during search and rescue tasks. However, under noisy conditions robotic implementations of microphone arrays might degrade their precision when localizing sound sources. For practical applications, human hearing still leaves behind microphone arrays. Daniel Kisch is an example of how humans are able to efficiently perform echo-localization to recognize their environment, even in noisy and reverberant environments. For ubiquitous computing, another limitation of acoustic localization algorithms is within their capabilities of performing real-time Digital Signal Processing (DSP) operations. To tackle those problems, tradeoffs between size, weight, cost and power consumption compromise the design of acoustic sensors for practical applications. This works presents the design and operation of a large microphone array for DSP applications in realistic environments. To address those problems this project introduces the Pyramic sound capture system designed at LAP in EPFL. Pyramic is a custom hardware which possesses 48 microphones distributed in the edges of a tetrahedron. The microphone arrays interact with a Terasic DE1-SoC board from Altera Cyclone V family devices, which combines a Hard Processor System (HPS) and a Field Programmable Gate Array (FPGA) in the same die. The HPS part integrates a dual-core ARM-based Cortex-A9 processor, which combined with the power of FPGA design suitable for processing multichannel microphone signals. This thesis explains the implementation of the Pyramic array. Moreover, FPGA-based hardware accelerators have been designed to implement a Master SPI communication with the array and a parallel 48 channels FIR filters cascade of the audio data for delay-and-sum beamforming applications. Additionally, the configuration of the HPS part allows the Pyramic array to be controlled through a Linux based OS. The main purpose of the project is to develop a flexible platform in which real-time echo-location algorithms can be implemented. The effectiveness of the Pyramic array design is illustrated by testing the recorded data with offline direction of arrival algorithms developed at LCAV in EPFL.

2016

ColTraIn: Co-located DNN training and inference

Mario Paulo Drumond Lages De Oliveira

Deep neural network inference accelerators are deployed at scale to accommodate online services, but face low average load because of service demand variability, leading to poor resource utilization. Unfortunately, reclaiming inference idle cycles is difficult, as no other workload can execute on such custom accelerators. DNN training services offer opportunities to reclaim inference accelerator idle cycles. However, the inference services' tight latency constraints and the training algorithms' dependence on floating-point arithmetic limit the opportunities for piggybacking training services on inference accelerators. In this thesis, we tackle the challenges that prevent inference DNN accelerators from exposing their idle cycles to training services. We first develop an efficient numeric representation that enables DNN training with accuracy similar to single-precision floating point and energy efficiency similar to 8-bit fixed point. Then, we explore the inference accelerator design space to show that, unlike in current latency-optimal platforms, relaxing latency constraints with ALU arrays that are batching-optimized achieves near-optimal throughput for a given area and power envelope. High throughput inference accelerators maximize the opportunities to piggyback training. Finally, we present Equinox, a family of inference accelerators designed to piggyback training. Equinox employs a uniform encoding and a priority hardware scheduler that processes training requests during inference idle cycles without affecting inference tail latency. Overall, we show that exposing accelerator idle cycles to training services uncovers significant computing power for training services with a small overhead for inference accelerators, improving overall datacenter efficiency.

EPFL2020

An Accelerator for High Efficient Vision Processing

Zidong Du, Robert Fasthuber, Qi Guo, Paolo Ienne

In recent years, neural network accelerators have been shown to achieve both high energy efficiency and high performance for a broad application scope within the important category of recognition and mining applications. Still, both the energy efficiency and performance of such accelerators remain limited by memory accesses. In this paper, we focus on image applications, arguably the most important category among recognition and mining applications. The neural networks which are state-of-the-art for these applications are convolutional neural networks (CNNs), and they have an important property: weights are shared among many neurons, considerably reducing the neural network memory footprint. This property allows to entirely map a CNN within an SRAM, eliminating all DRAM accesses for weights. By further hoisting this accelerator next to the image sensor, it is possible to eliminate all remaining DRAM accesses, i.e., for inputs and outputs. In this paper, we propose such a CNN accelerator, placed next to a CMOS or CCD sensor. The absence of DRAM accesses combined with a careful exploitation of the specific data access patterns within CNNs allows us to design an accelerator which is highly energy-efficient. We present a single-core implementation down to the layout at 65 nm, with a modest footprint of 5.94mm(2) and consuming only 336mm(2), but still about 30x faster than high-end GPUs. For visual processing with higher resolution and frame-rate requirements, we further present a multicore implementation with elevated performance.

Ieee-Inst Electrical Electronics Engineers Inc2017

Pyramic array: An FPGA based platform for multi-channel audio acquisition

Juan Azcarreta Ortiz

2016