I was reading about a new brain-inspired chip on the BBC, which suggests new potentially cost effective ways of executing heavy data/compute tasks. They speak of “neurons” and “synapses” as a way to describe it, which is interesting as we can use humans as a reference for what we might want to achieve.
Wikipedia suggests that the human brain holds 86 billion neurons and that each neuron has on average 7,000 synaptic connections to other neurons.
Let’s play with the idea of x.ai using an 8 byte floating point number for each of those connections – as we build an AI network to hold each one of these in memory. The fantasy of making Amy self-aware.
RAM: 86,000,000,000 * 7,000 * 8 bytes = 4.8 Petabytes
A memory optimized machine on Amazon AWS, with 244 Gigabyte of memory and a little compute power to manage it will cost you $2.8 per hour.
Money: (4.8 Petabytes / 244 Gigabytes) * 24h * 30d * $2.8
= $39,659,016 per month (or $476 Million per year)
Going back to the article. Dr Modha suggested that the chip is “endlessly scalable” and that this isn’t a 10-15% improvement. He said. “You’re talking about orders and orders of magnitude.”. One would hope so, because we only invested ~$2M in our seed round 😉