Anybody did run the numbers? Not to mention the recent discoveries that the human neuron/synapse encode way more states than expected.
I think it's pretty clear that an entire human brain is not required for the operation of consciousness. Certain people have lost massive portions of their brain and still maintained regular consciousness functioning. On top of that, the typical human brain only consumes about 20 watts of power to do its thing. So that's just a safe upper bound for the power required.
Over the last decade we've seen the rise of ML systems that have replicated or surpassed capabilities long thought to be the exclusive domain of the human brain. Think facial recognition, AlphaGo or the very recent DALL-E 2.
This had led me to the personal belief that we (in aggregate) likely already have the computing power to achieve not only artificial consciousness, but also AGI and beyond. We simply haven't figured out the correct model connectivity and parameters.
So while we undoubtedly have sufficient aggregate computing power, I don't think it is structured in a way that would allow it to emulate human cognition and problem sovling, not just pattern recognition.
That said, consciousness is an incredibly heterogeneously defined term but can probably be pretty widely applied (many people confuse consciousness with sentience or sapience.)
Neurons are all giving each-other feedback in real time, which leads to emergent behavior. To really reproduce this, you would probably need a very different hardware paradigm.
So if we wanted to make a simulation with reasonable fidelity to the mechanisms of cellular information processing we know about, I would wager to guess it's more on the order of 1GPU == 1Synapse at a minimum.
That's not to say that all of this complexity is required to reproduce consciousness in a meaningful way, but actual information processing in the brain is gob-smackingly complex.
The interesting thing here is, what V1 does is it computes directional receptive fields based on raw data from the retinas. We can implement that in a tiny silicon chip, maybe even just a DSP chip, not a full CPU. We know how to implement this kind of computation super efficiently, and without using neural networks.
We don't fully understand what goes on in many areas of the brain, but if we can find efficient ways to implement equivalent computations, we might technically already have sufficient manufacturing technology to implement equivalent functionality in a pretty compact and power-efficient form factor, or we might not be that far from there.
A classic quote from Dijkstra: "The question of whether machines can think is about as relevant as the question of whether submarines can swim."
If there were a brain region that might be emulated to reproduce "general information processing", the best candidate would probably be neocortex. As far as I am aware, that does seem to be a very wide region of the brain which has fairly consistent structure, and largely serves to integrate all other systems.
The point is that V1 is a large part of the visual cortex. It's something the brain dedicates a lot of resources to, and we could implement something similar in silicon almost trivially. Your smartphone GPU may have enough compute to simulate an equivalent computation at 200 frames per second.
It's possible that in the future, we'll be able to simulate consciousness very efficiently once we have a better handle on the kinds of computations that are involved, much more efficiently than what the brain does.
But even in this noisy environment, our brains can usually operate just fine. Given this level of background noise, I would be very surprised if the average synapse would require more than a 16-bit floating point value for its outputs.
Some synapses contain multiple receptor sub-types. Dopamine, for example, has some receptors which are inhibitory, and others which are excitatory. Each receptor is sensitive to different concentrations of dopamine. This gives the effect that when the dopamine concentrations in the synapse are low, this synapse acts to inhibit the activity of the efferent neuron, while when the concentration reaches a certain threshold, the output of the synapse flips to an excitatory signal.
How would you model that system with a single F16?
And that's just one example: for instance there are specific signals which trigger cascading effects in the neuron which up-regulate certain chemical pathways, causing an amplifying effect on other signals.
Or you have shunting effects, where an inhibitory synapse placed at a certain point in the dendritic arbor serves to selectively cancel signals from farther down that one specific branch of the dendrite.
That's not "noise" - the brain has incredibly refined and detailed mechanisms for information processing which go very far beyond `weight * sigmoid(activation) = output`