Scaling of standard deviation of noise in memristor weights

[HeatPhoenix]

I'm trying to properly understand the effect of noise within the simulation (and within memristor-based accelerators in general).

Looking at this image above, and my own tests below. When reading the weights, either with a normalized vector or during a multiplication, it seems to me that the variation around the real weight isn't multiplicative, but mostly additive.

In this graph, I am sweeping what the read-out result is when I set all the weights to a particular value, and then (realistically) read them out repeatedly to discover the distribution. Here's a small sample of the code I used:

weights, bias = model.get_weights()
weights.fill_(value)
model.set_weights(weights, bias)

weights_0_list = []

# Run 1000 times and get weights[0]
for _ in range(1000):
    weights, bias = model.get_weights(realistic=True)
    weights_0_list.append(weights[0].cpu().numpy()) 

So the key question here is: At a set weight of 10, compared to a set weight 1 of 0. 1, the std dev doesn't seem to change proportionally. Does this make sense?

And if not, what am I misunderstanding/doing wrong? I would have expected it to be 5 times as large at 10 compared to 2, for example.