Detailed brain modeling has been presenting significant challenges to the world of high-performance computing (HPC), posing computational problems that can benefit from modern hardware-acceleration technologies. We explore the capacity of GPUs for simulating large-scale neuronal networks based on the Adaptive Exponential neuron-model, which is widely used in the neuroscientific community. Our GPU-powered simulator acts as a benchmark to evaluate the strengths and limitations of modern GPUs, as well as to explore their scaling properties when simulating large neural networks. This work presents an optimized GPU implementation that outperforms a reference multicore implementation by 50x, whereas utilizing a dual-GPU configuration can deliver a speedup of 90x for networks of 20,000 fully interconnected AdEx neurons.

Mathematical models with varying degrees of complexity have been proposed and simulated in an attempt to represent the intricate mechanisms of the human neuron. One of the most biochemically realistic and analytical models, based on the Hodgkin–Huxley (HH) model, has been selected for study in this paper. In order to satisfy the model's computational demands, we present a simulator implemented on Intel Xeon Phi Knights Landing manycore processors. This high-performance platform features an x86-based architecture, allowing our implementation to be portable to other common manycore processing machines. This is reinforced by the fact that Phi adopts the popular OpenMP and MPI programming models. The simulator performance is evaluated when calculating neuronal networks of varying sizes, density and network connectivity maps. The evaluation leads to an analysis of the neuronal synaptic patterns and their impact on performance when tackling this type of workload on a multinode system. It will be shown that the simulator can calculate 100 ms of simulated brain activity for up to 2 millions of biophysically-accurate neurons and 2 billion neuronal synapses within one minute of execution time. This level of performance renders the application an efficient solution for large-scale detailed model simulation.

Emerging cloud applications like machine learning, AI, big data analytics and scientific computing require highperformance computing systems that can sustain the increased amount of data processing without consuming excessive power. To this end, many cloud operators have started deploying hardware accelerators, like GPUs and FPGAs, to increase the performance of computationally intensive tasks. However, increased performance, comes at a higher cost of increased programming complexity for utilizing these accelerators. VINEYARD has developed a versatile framework that allows the seamless deployment and utilization of heterogeneous accelerators in the cloud without increasing the programming complexity while offering the flexibility of software packages. This paper presents the main components that have been developed in the VINEYARD framework and focuses on BrainFrame, the neurocomputing case that demonstrates the new framework's value. BrainFrame not only accelerates neuronal simulations but also has an architecture that allows easy access to neuroscientists, hiding the system complexity, and enabling a modular integration of new accelerated simulators.

Biologically accurate neuron simulations are increasingly important in research related to brain activity. They are computationally intensive and feature data and task parallelism. In this paper, we present a case study for the mapping of a biologically accurate inferior-olive (InfOli), neural cell simulator on an many-core research platform. The Single-Chip Cloud Computer (SCC) is an experimental processor created by Intel Labs. The target neurons provide a major input to the cerebellum and are involved in motor skills and space perception. We exploit task-and data-partitioning, scaling the simulation over more than 40,000 neurons. The voltage-and frequency-scaling capabilities of the chip are explored, achieving more than 20% energy savings with negligible performance degradation. Four platform configurations are evaluated and a mapping with balanced workload and constant voltage and frequency is formally derived as optimal.