I’m sure somebody has made these kinds of analyses in much greater detail, but I wanted to get a sense of the computational limits that we are presently at. Caution: Many of the following numbers are pulled out of the internet without serious effort.

## Energy efficiency

At its developed peak, the brain might have 200 trillion synapses and consume 1760 kJ/day. Making this a ratio, we get 200 trillion synapses / 20W = 10 trillion synapses / watt. Suppose 1 synapse performs at least 10 ‘operations’ per second. Then the brain performs at least 100 TFLOPS with an ideal efficiency exceeding 100 TFLOPS/watt. Compare this to machine computation:

• This paper (fig 5) says that in 2020, GPUs reach 100 GFLOPS/watt. However it notes that energy efficiency is exponentially increasing (rough estimate, 10x every 10 years)
• The v3-32 TPU Pod delivers ~1680 TFLOPS (see below) with estimated power consumption 200W/core*32cores = 6.4kW. This makes about 250GFLOPS/watt. Only about 3 orders of magnitude less than this brain estimate.