Comparison of Spiking Neural Network Simulator Performance. See our pre-print presenting these results in detail; Spike: A GPU Optimised Spiking Neural Network Simulator
Simulators are included as git submodules. Currently the simulators being compared are:
In order to clone the repository AND all of the specific simulator versions used, please run;
git clone --recurse-submodules https://github.com/nasiryahm/SNNSimulatorComparison
Source Paper: Vogels TP, Abbott LF. Signal propagation and logic gating in networks of integrate-and-fire neurons. J Neurosci. 2005;25: 10786–10795.
Simulations were all single-threaded and run on machine with: CPU: Intel i7-4770K GPU: NVIDIA GTX 1070 founders edition OS: Ubuntu 16.04 LTS
All simulations load the same set of weights (ee/ei/ie/ii.wmat which are expected to be located in the Benchmark folders) and therefore allow a direct simulation comparison. This benchmark simulation was inspired by the benchmarks used for the Auryn simulator. See their github repository for further details.
The installed python (conda) environment details are located in the conda_req.txt file.
Results of a simulation of the VogelsAbbott Benchmark with synaptic delays either all set to 0.1ms, or 0.8ms (with a simulation timestep of 0.1ms in both cases).
Note that all simulators other than NEST employ a forward euler solver to compute updates to the network dynamics, hench NEST is shown in gray.
A comparison of the ISI distributions, firing rasters, and firing rates is present in an iPython notebook. These results were produced from files which are automatically dumped when the "--fast" option is not used in simulation execution.
Above, only Spike and Auryn are compared. Auryn is benchmarked with 1, 2, 4, and 8 threads on a system with a 16 core Intel Xeon E5-2623 v4. These benchmarks are shown as black points on the plot above. An exponential decay curve is fit to the Auryn datapoints as shown in black. For comparison, the single-threaded, single-GPU speed of Spike is shown in red.
Source Paper: Brunel N. Dynamics of sparsely connected networks of excitatory and inhibitory spiking neurons. J Comput Neurosci. 2000;8: 183–208.
The model implemented is a specific version of this network as implemented by the Auryn and ANNarchy networks (see repositories). The connectivity for this network produces files which are too large to be contained within this repository. Before running the Brunel benchmarks, first run the executable createConnectivity in order to produce a set of connectivity matrices using Auryn. This requires that Auryn is installed in the Simulators/auryn directory.
Results of a simulation of the Brunel10K Benchmark with and without synaptic plasticity
Note that Brian2, NEST, and Auryn use input stimulation methods which approximate the effect of Poisson Firing Input Neurons in order to achieve a speedup. Spike, GeNN and ANNarchy provide inputs through modelled neurons with Poisson distribution sampled spike times.
Above, only Spike and Auryn are compared. Auryn is benchmarked with 1, 2, 4, and 8 threads on a system with a 16 core Intel Xeon E5-2623 v4. These benchmarks are shown as black points on the plot above. An exponential decay curve is fit to the Auryn datapoints as shown in black. For comparison, the single-threaded, single-GPU speed of Spike is shown in red.
Spike, Auryn and NEST simulator are auto compiled (using make). Ensure that the dependencies for these libraries are pre-installed. To see these, please visit the github pages for these projects.
ANNarchy and Brian can be installed using pip install Simulators/ANNarchy/ and pip install Simulators/brian2 from this location. Again, ensure that the dependencies are pre-installed.
The dependencies and packages required for the installation of these two packages can be found in the requirements.txt file in the root of this repo.
After installing/compiling the simulators (located in the Simulators folder), the models can be compiled as necessary in the Benchmarks folder. Examples of how to compile Spike, Auryn, and GeNN models are given as "compile.sh" bash scripts. Each model has a few key command line arguments which can be used in order to set the behaviour;
Setting the number of seconds (X) for which the simulation will run;
--simtime X
A boolean which will ensure that any spike recording is turned off and the model execution is timed;
--fast
Spike, Brian2, and NEST simulator support ranges of delays.
ANNarchy informs us that it can handle uniform delays (experimentally) though does not compile in this case. Auryn and GeNN do not currently support ranges of delays within a single synaptic population.