Spice (/spaɪk/) is a multi-GPU, time-driven (aka clock-based), general-purpose spiking neural network simulator. Its features include:
- State of the art performance (incl. sub-second setup)
- Multi-GPU support with linear scaling up to 8 GPUs
- Ability to define custom models in native C++
- Modern, user-friendly API
- Even Faster SNN Simulation with Lazy+Event-driven Plasticity and Shared Atomics (2021) (branch:gather)
- Multi-GPU SNN Simulation with Perfect Static Load Balancing (2021)
- Faster and Simpler SNN Simulation with Work Queues (2020)
- CUDA SDK 11 or higher
- nVidia GPU with compute capability 6.1 (10XX or newer)
- CMAKE 3.0.2 or higher, recommended default settings, builds out of the box.
- C++17 compatible compiler
- (Only tested on Linux: Spice contains no Linux-specific code and should build on Windows, but hasn't been tested on it for some time.)
The "samples" project currently implements benchmarks used for our publications. It compiles to an executable that can be invoked via
./samples --model {vogels|brunel|brunel+} --nsyn 1234 --gpu {single|multi} --bench {sim|setup}
which prints a json object similar to
{
"sim": "samples",
"model": "brunel",
"#syn": 1000000000,
"#gpus": 1,
"simtime": 1.49136,
"setuptime": 0.081864
}
The field "simtime" is the ratio between simulation time (wall clock time) and biological time. All benchmarks simulate 10s of biological time. "setuptime" is the absolute setup time in seconds.
Have a look at the sample models defined in spice/models; the syntax is pretty sraight-forward. Currently, the easiest way to define your own models is to hack one of the existing ones. To instantiate your model you'd write
cuda::snn<mymodel> net(
{10, 20, 30}, // create three neuron populations with 10, 20, 30 neurons respectively
{{0, 2, 0.1}, // randomly connect pop. A->C with prob. 10%
{1, 2, 0.05}}, // randomly connect pop. B->C with prob. 5%
0.001, // time step in seconds
8 // synaptic delay
);
while (true)
net.step();
cuda::snn can be any of {cpu::snn|cuda::snn|cuda::multi_snn} depending on whether you want to run on a (single core) CPU, single GPU, or multiple GPUs. Beware that net.step() executes delay many steps when net is of type cuda::multi_snn in order to be able to hide latency from spike synchronization. net.step() also takes an optional pointer to a std::vector<int> and writes spiking data into it.