Dreaming of a Zettaflop: Harnessing Unimaginable Computing Power

As the dawn of unprecedented computing power approaches, I find myself both excited and unprepared. The vision of owning a zettaflop machine—a system capable of performing 1e21 floating-point operations per second—drives my ambition. While current technology like comma's exaflop is impressive, I dream of scaling up to a zettaflop, which would allow for unimaginable feats such as processing 1e27 training runs or compressing a century's worth of work into mere weeks.

The sensory overload of such power is akin to experiencing a vast torrent of data, as described by Vernor Vinge in 'True Names'. I yearn to command this power, much like I do with my teraflop laptop or petaflop systems. The potential of a zettaflop machine is staggering: it could search every book in history, solve complex problems, and iterate code to perfection, essentially having 50,000 minds working in unison.

However, the biggest hurdle remains power consumption. Current technology, like B200s achieving 10 TFLOPS/W, requires significant energy, and I estimate needing 10 MW. Solar power could be a solution, with 250 acres of panels providing the necessary energy. The financial investment is substantial—$10 million each for the machine, solar panels, and land—but I am determined to achieve this dream within my lifetime.