I already have some implementation of the parallelization in algorithm of SOM.
Let us quickly remind the main steps of the classical Kohonen learning algorithm.

1. Initialization – means setting values to the network, either random or from input space.
2. Sampling – select random vector x from the input space.
3. Similarity matching – find best matching neuron, i.e. neuron which is most similar to the x.
4. Update – need to update neighbourhood neurons to the BMN in particular radius r, which is decreasing with time.
5. Continuation (2-5) – we repeat 2-5 many times, while do not have map with enough quality.

As you see algorithm could be very expensive if we have big amounts of data.

Also, steps 3 and 4 takes the most of time, what if we execute 2-3-5 in separate threads? Yes, we could do this to some extend. Main issue is when we have overlapping of affected area by two best matched neurons wich we found in separate threads.

I’m bit tired to write a lot of explanations of this algorithm, so I prepared 3 images that says for themselves. Hope will have detailed explanations and different tuning things for this soon.