= 27d567bc2d48d9f40e9ccc20ea370b15)
V_V comments on MIRI's Approach - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (59)
I don't think that "switches" per second is a relevant metric here. The computation performed by a single neuron in a single firing cycle is much more complex than the computation performed by a logic gate in a single switching cycle.
The amount of computational power required to simulate a human brain in real time is estimated in the petaflops range. Only the largest supercomputer operate in that range, certainly not common GPUs.
You misunderstood me - the biological switch events I was referring to are synaptic ops, and they are comparable to transistor/gate switch ops in terms of minimum fundemental energy cost in Landauer analysis.
That is a tad too high, the more accurate figure is 10^14 ops/second (10^14 synapses * avg 1 hz spike rate). The minimal computation required to simulate a single GPU in real time is 10,000 times higher.
I've seen various people give estimates in the order of 10^16 flops by considering the maximum firing rate of a typical neuron (~10^2 Hz) rather than the average firing rate, as you do.
On one hand, a neuron must do some computation whether it fires or not, and a "naive" simulation would necessarily use a cycle frequency of the order of 10^2 Hz or more, on the other hand, if the result of a computation is almost always "do not fire", then as a random variable the result has little information entropy and this may perhaps be exploited to optimize the computation. I don't have a strong intuition about this.
On a traditional CPU perhaps, on another GPU I don't think so.