SearcharxivSearch

arXiv subjects

Ottorino Frezza

Publications and source records attributed to Ottorino Frezza.

17 recordsLinked to original sources

AIGOR: A Modular, Event-Driven Neuromorphic Architecture for Configurable SNN Inference

Spiking neural networks (SNNs) run today on a fragmented landscape of hardware: dedicated neuromorphic processors, application-specific FPGA accelerators, and large-scale neuroscience simulators, each typically built around a fixed neuron model, execution strategy, or workload class. We present AIGOR, a modular, event-driven neuromorphic architecture for spiking neural network inference. AIGOR organizes neurons into timestep-synchronized processing cores that exchange spikes as packets over a packet-switched communication layer, and it is assembled from a library of parameterized compute, memory, and communication IP blocks rather than as a one-off design for a single network. The neuron model, numeric precision, the folding of neurons onto hardware, and the partitioning across cores are configured per instance rather than committed at design time; a single declarative specification then generates the cores, neuron kernels, and synaptic-memory images that realize a given network. We validate a first prototype on the AMD Versal VPK180 across two deliberately different workloads mapped onto the same cores: a feedforward image classifier trained in snnTorch and a recurrent bal anced random network modeled in NEST. The classifier reproduces its snnTorch reference accuracy, and the recurrent network matches its NEST reference at spike-level precision across multiple cores spanning two FPGAs. We report post-implementation resource utilization and validate the multi-node synchronization scheme in simulation up to one thousand cores on a three-dimensional torus. The prototype's measured limits localize the throughput bottleneck in the synaptic-delivery datapath and the global timestep barrier, and motivate a set of datapath refinements, now in development, that the configurable structure of the architecture admits as changes to the same cores.

cs.AR

Online Data Reduction with Spiking Neural Networks: A Temporal-Coincidence Encoder and Distributed SNN for the ePIC dRICH Detector

The dual-radiator Ring Imaging Cherenkov (dRICH) detector of the ePIC experiment at the Electron-Ion Collider (EIC) will read out $\sim$320,000 silicon photomultiplier (SiPM) channels at a bunch-crossing rate of 100 MHz. The dark count rate (DCR) of the SiPMs is expected to rise up to 300 kHz per channel over the experiment lifetime, saturating the output bandwidth and requiring an online data reduction factor of at least five. Most crossings contain only uncorrelated DCR hits, while genuine Cherenkov hits cluster within $\sim$2 ns of the 10 ns crossing window: an intrinsically temporal discrimination problem. We present a two-stage online data reduction pipeline based on spiking neural networks (SNNs). The first stage is a per-photodetection-unit leaky-integrate-and-fire (LIF) temporal coincidence encoder that converts raw SiPM hits into a sparse spike stream, achieving over 90% data sparsification before any learned classifier is applied. The second stage is a distributed SNN (30 sub-sector networks plus an aggregation network) deployed on FELIX-155 DAM boards and a dedicated Trigger Processor board, classifying each crossing as Noise-Only or Signal+Noise. On simulated ePIC events the system reaches a true positive rate above 94% at a true negative rate of at least 80% across the full DCR range; an optional early-exit strategy reduces the average classification latency to $\sim$2 algorithmic timesteps at the cost of a few percentage points on both metrics. A hardware proof-of-concept on an AMD Versal Premium FPGA, integrating the LIF encoder with the AIGOR multi-core neuromorphic architecture, validates a single sub-sector pipeline at $\sim$1.7 MHz throughput; ongoing work targets 100 MHz through an identified set of optimizations of the inference fabric. The methodology may be relevant to other timing-driven detector applications at high rate.

physics.ins-det

APEIRON: composing smart TDAQ systems for high energy physics experiments

We present APEIRON, a distributed heterogeneous processing framework comprising both hardware architecture and software stack for multi-FPGA systems. Targeting smart trigger and data acquisition (TDAQ) systems in high energy physics, APEIRON spans the full software hierarchy: from low-level device drivers to a high-level dataflow programming model based on High-Level Synthesis. We describe the framework design, its core communication infrastructure, and a particle identification application for the NA62 experiment as a representative physics use case.

physics.ins-det

NET4EXA: Pioneering the Future of Interconnects for Supercomputing and AI

NET4EXA aims to develop a next-generation high-performance interconnect for HPC and AI systems, addressing the increasing demands of large-scale infrastructures, such as those required for training Large Language Models. Building upon the proven BXI (Bull eXascale Interconnect) European technology used in TOP15 supercomputers, NET4EXA will deliver the new BXI release, BXIv3, a complete hardware and software interconnect solution, including switch and network interface components. The project will integrate a fully functional pilot system at TRL 8, ready for deployment into upcoming exascale and post-exascale systems from 2025 onward. Leveraging prior research from European initiatives like RED-SEA, the previous achievements of consortium partners and over 20 years of expertise from BULL, NET4EXA also lays the groundwork for the future generation of BXI, BXIv4, providing analysis and preliminary design. The project will use a hybrid development and co-design approach, combining commercial switch technology with custom IP and FPGA-based NICs. Performances of NET4EXA BXIv3 interconnect will be evaluated using a broad portfolio of benchmarks, scientific scalable applications, and AI workloads.

cs.NI

APEIRON: composing smart TDAQ systems for high energy physics experiments

APEIRON is a framework encompassing the general architecture of a distributed heterogeneous processing platform and the corresponding software stack, from the low level device drivers up to the high level programming model. The framework is designed to be efficiently used for studying, prototyping and deploying smart trigger and data acquisition (TDAQ) systems for high energy physics experiments.

cs.DC

Impact of exponential long range and Gaussian short range lateral connectivity on the distributed simulation of neural networks including up to 30 billion synapses

Recent experimental neuroscience studies are pointing out the role of long-range intra-areal connectivity that can be modeled by a distance dependent exponential decay of the synaptic probability distribution. This short report provides a preliminary measure of the impact of exponentially decaying lateral connectivity compared to that of shorter-range Gaussian decays on the scaling behaviour and memory occupation of a distributed spiking neural network simulator (DPSNN). Two-dimensional grids of cortical columns composed by point-like spiking neurons have been connected by up to 30 billion synapses using exponential and Gaussian connectivity models. Up to 1024 hardware cores, hosted on a 64 nodes server platform, executed the MPI processes composing the distributed simulator. The hardware platform was a cluster of IBM NX360 M5 16-core compute nodes, each one containing two Intel Xeon Haswell 8-core E5-2630 v3 processors, with a clock of 2.40GHz, interconnected through an InfiniBand network. This study is conducted in the framework of the CORTICONIC FET project, also in view of the next -to-start activities foreseen as part of the Human Brain Project (HBP), SubProject 3 Cognitive and Systems Neuroscience, WaveScalES work-package.

cs.DC

Scaling to 1024 software processes and hardware cores of the distributed simulation of a spiking neural network including up to 20G synapses

This short report describes the scaling, up to 1024 software processes and hardware cores, of a distributed simulator of plastic spiking neural networks. A previous report demonstrated good scalability of the simulator up to 128 processes. Herein we extend the speed-up measurements and strong and weak scaling analysis of the simulator to the range between 1 and 1024 software processes and hardware cores. We simulated two-dimensional grids of cortical columns including up to ~20G synapses connecting ~11M neurons. The neural network was distributed over a set of MPI processes and the simulations were run on a server platform composed of up to 64 dual-socket nodes, each socket equipped with Intel Haswell E5-2630 v3 processors (8 cores @ 2.4 GHz clock). All nodes are interconned through an InfiniBand network. The DPSNN simulator has been developed by INFN in the framework of EURETILE and CORTICONIC European FET Project and will be used by the WaveScalEW tem in the framework of the Human Brain Project (HBP), SubProject 2 - Cognitive and Systems Neuroscience. This report lays the groundwork for a more thorough comparison with the neural simulation tool NEST.

cs.DC

Power, Energy and Speed of Embedded and Server Multi-Cores applied to Distributed Simulation of Spiking Neural Networks: ARM in NVIDIA Tegra vs Intel Xeon quad-cores

This short note regards a comparison of instantaneous power, total energy consumption, execution time and energetic cost per synaptic event of a spiking neural network simulator (DPSNN-STDP) distributed on MPI processes when executed either on an embedded platform (based on a dual socket quad-core ARM platform) or a server platform (INTEL-based quad-core dual socket platform). We also compare the measure with those reported by leading custom and semi-custom designs: TrueNorth and SpiNNaker. In summary, we observed that: 1- we spent 2.2 micro-Joule per simulated event on the "embedded platform", approx. 4.4 times lower than what was spent by the "server platform"; 2- the instantaneous power consumption of the "embedded platform" was 14.4 times better than the "server" one; 3- the server platform is a factor 3.3 faster. The "embedded platform" is made of NVIDIA Jetson TK1 boards, interconnected by Ethernet, each mounting a Tegra K1 chip including a quad-core ARM Cortex-A15 at 2.3GHz. The "server platform" is based on dual-socket quad-core Intel Xeon CPUs (E5620 at 2.4GHz). The measures were obtained with the DPSNN-STDP simulator (Distributed Simulator of Polychronous Spiking Neural Network with synaptic Spike Timing Dependent Plasticity) developed by INFN, that already proved its efficient scalability and execution speed-up on hundreds of similar "server" cores and MPI processes, applied to neural nets composed of several billions of synapses.

cs.DC

EURETILE D7.3 - Dynamic DAL benchmark coding, measurements on MPI version of DPSNN-STDP (distributed plastic spiking neural net) and improvements to other DAL codes

The EURETILE project required the selection and coding of a set of dedicated benchmarks. The project is about the software and hardware architecture of future many-tile distributed fault-tolerant systems. We focus on dynamic workloads characterised by heavy numerical processing requirements. The ambition is to identify common techniques that could be applied to both the Embedded Systems and HPC domains. This document is the first public deliverable of Work Package 7: Challenging Tiled Applications.

cs.DC

Distributed simulation of polychronous and plastic spiking neural networks: strong and weak scaling of a representative mini-application benchmark executed on a small-scale commodity cluster

We introduce a natively distributed mini-application benchmark representative of plastic spiking neural network simulators. It can be used to measure performances of existing computing platforms and to drive the development of future parallel/distributed computing systems dedicated to the simulation of plastic spiking networks. The mini-application is designed to generate spiking behaviors and synaptic connectivity that do not change when the number of hardware processing nodes is varied, simplifying the quantitative study of scalability on commodity and custom architectures. Here, we present the strong and weak scaling and the profiling of the computational/communication components of the DPSNN-STDP benchmark (Distributed Simulation of Polychronous Spiking Neural Network with synaptic Spike-Timing Dependent Plasticity). In this first test, we used the benchmark to exercise a small-scale cluster of commodity processors (varying the number of used physical cores from 1 to 128). The cluster was interconnected through a commodity network. Bidimensional grids of columns composed of Izhikevich neurons projected synapses locally and toward first, second and third neighboring columns. The size of the simulated network varied from 6.6 Giga synapses down to 200 K synapses. The code demonstrated to be fast and scalable: 10 wall clock seconds were required to simulate one second of activity and plasticity (per Hertz of average firing rate) of a network composed by 3.2 G synapses running on 128 hardware cores clocked @ 2.4 GHz. The mini-application has been designed to be easily interfaced with standard and custom software and hardware communication interfaces. It has been designed from its foundation to be natively distributed and parallel, and should not pose major obstacles against distribution and parallelization on several platforms.

cs.DC

NaNet:a low-latency NIC enabling GPU-based, real-time low level trigger systems

We implemented the NaNet FPGA-based PCI2 Gen2 GbE/APElink NIC, featuring GPUDirect RDMA capabilities and UDP protocol management offloading. NaNet is able to receive a UDP input data stream from its GbE interface and redirect it, without any intermediate buffering or CPU intervention, to the memory of a Fermi/Kepler GPU hosted on the same PCIe bus, provided that the two devices share the same upstream root complex. Synthetic benchmarks for latency and bandwidth are presented. We describe how NaNet can be employed in the prototype of the GPU-based RICH low-level trigger processor of the NA62 CERN experiment, to implement the data link between the TEL62 readout boards and the low level trigger processor. Results for the throughput and latency of the integrated system are presented and discussed.

physics.ins-det

Architectural improvements and 28 nm FPGA implementation of the APEnet+ 3D Torus network for hybrid HPC systems

Modern Graphics Processing Units (GPUs) are now considered accelerators for general purpose computation. A tight interaction between the GPU and the interconnection network is the strategy to express the full potential on capability computing of a multi-GPU system on large HPC clusters; that is the reason why an efficient and scalable interconnect is a key technology to finally deliver GPUs for scientific HPC. In this paper we show the latest architectural and performance improvement of the APEnet+ network fabric, a FPGA-based PCIe board with 6 fully bidirectional off-board links with 34 Gbps of raw bandwidth per direction, and X8 Gen2 bandwidth towards the host PC. The board implements a Remote Direct Memory Access (RDMA) protocol that leverages upon peer-to-peer (P2P) capabilities of Fermi- and Kepler-class NVIDIA GPUs to obtain real zero-copy, low-latency GPU-to-GPU transfers. Finally, we report on the development activities for 2013 focusing on the adoption of the latest generation 28 nm FPGAs and the preliminary tests performed on this new platform.

cs.AR

GPU peer-to-peer techniques applied to a cluster interconnect

Modern GPUs support special protocols to exchange data directly across the PCI Express bus. While these protocols could be used to reduce GPU data transmission times, basically by avoiding staging to host memory, they require specific hardware features which are not available on current generation network adapters. In this paper we describe the architectural modifications required to implement peer-to-peer access to NVIDIA Fermi- and Kepler-class GPUs on an FPGA-based cluster interconnect. Besides, the current software implementation, which integrates this feature by minimally extending the RDMA programming model, is discussed, as well as some issues raised while employing it in a higher level API like MPI. Finally, the current limits of the technique are studied by analyzing the performance improvements on low-level benchmarks and on two GPU-accelerated applications, showing when and how they seem to benefit from the GPU peer-to-peer method.

physics.comp-ph

A heterogeneous many-core platform for experiments on scalable custom interconnects and management of fault and critical events, applied to many-process applications: Vol. II, 2012 technical report

This is the second of a planned collection of four yearly volumes describing the deployment of a heterogeneous many-core platform for experiments on scalable custom interconnects and management of fault and critical events, applied to many-process applications. This volume covers several topics, among which: 1- a system for awareness of faults and critical events (named LO|FA|MO) on experimental heterogeneous many-core hardware platforms; 2- the integration and test of the experimental hardware heterogeneous many-core platform QUoNG, based on the APEnet+ custom interconnect; 3- the design of a Software-Programmable Distributed Network Processor architecture (DNP) using ASIP technology; 4- the initial stages of design of a new DNP generation onto a 28nm FPGA. These developments were performed in the framework of the EURETILE European Project under the Grant Agreement no. 247846.

cs.DC

'Mutual Watch-dog Networking': Distributed Awareness of Faults and Critical Events in Petascale/Exascale systems

Many tile systems require techniques to be applied to increase components resilience and control the FIT (Failures In Time) rate. When scaling to peta- exa-scale systems the FIT rate may become unacceptable due to component numerosity, requiring more systemic countermeasures. Thus, the ability to be fault aware, i.e. to detect and collect information about fault and critical events, is a necessary feature that large scale distributed architectures must provide in order to apply systemic fault tolerance techniques. In this context, the LO|FA|MO approach is a way to obtain systemic fault awareness, by implementing a mutual watchdog mechanism and guaranteeing fault detection in a no-single-point-of-failure fashion. This document contains specification and implementation details about this approach, in the shape of a technical report.

cs.DC

APEnet+: high bandwidth 3D torus direct network for petaflops scale commodity clusters

We describe herein the APElink+ board, a PCIe interconnect adapter featuring the latest advances in wire speed and interface technology plus hardware support for a RDMA programming model and experimental acceleration of GPU networking; this design allows us to build a low latency, high bandwidth PC cluster, the APEnet+ network, the new generation of our cost-effective, tens-of-thousands-scalable cluster network architecture. Some test results and characterization of data transmission of a complete testbench, based on a commercial development card mounting an Altera FPGA, are provided.

physics.comp-ph

APEnet+: a 3D toroidal network enabling Petaflops scale Lattice QCD simulations on commodity clusters

Many scientific computations need multi-node parallelism for matching up both space (memory) and time (speed) ever-increasing requirements. The use of GPUs as accelerators introduces yet another level of complexity for the programmer and may potentially result in large overheads due to the complex memory hierarchy. Additionally, top-notch problems may easily employ more than a Petaflops of sustained computing power, requiring thousands of GPUs orchestrated with some parallel programming model. Here we describe APEnet+, the new generation of our interconnect, which scales up to tens of thousands of nodes with linear cost, thus improving the price/performance ratio on large clusters. The project target is the development of the Apelink+ host adapter featuring a low latency, high bandwidth direct network, state-of-the-art wire speeds on the links and a PCIe X8 gen2 host interface. It features hardware support for the RDMA programming model and experimental acceleration of GPU networking. A Linux kernel driver, a set of low-level RDMA APIs and an OpenMPI library driver are available, allowing for painless porting of standard applications. Finally, we give an insight of future work and intended developments.

hep-lat