search

nasiryahm / SNNSimulatorComparison

Comparison of Spiking Neural Network Simulator Performance
20 stars 9 forks source link

Spike:Brunel_w_plasticity slow #3

Open nkyriazis opened 5 years ago

nkyriazis commented 5 years ago

In the benchmarks, I see that in general (VA and Brunel) Spike is faster than Genn. However, in the brunel w plasticity case, Spike is an order of magnitude slower than Genn. It makes me think that plasticity might be computed on the CPU rather than on the GPU. Is this possible? How can I tell? Is this expected?

nasiryahm commented 5 years ago

Hi Nikolaos, are you comparing Spike with plasticity on against GeNN with plasticity off? In the simulation results on the homepage here, if you compare apples to apples (i.e. comparing only the red bars), Spike is faster than GeNN when both have plasticity on and (comparing only black bars) Spike is faster than GeNN when they both have plasticity off. The main issue is that when plasticity is switched on, both simulators (in fact almost all simulators) experience a slow down of an order of magnitude. Hopefully that clarifies. But if I misunderstood, I'm all ears.

nasiryahm commented 5 years ago

Oh also, I should mention that all plasticity for both the GeNN and Spike simulators is calculated on the GPU.