The performance capabilities of computers (e.g. CPUs, GPUs and Supercomputers) are expressed in floating point operations per second (FLOPS), a standard rate for indicating the number of floating-point arithmetic calculations systems can perform per second.

Currently, the NVIDIA TITAN V GPU, has one of the lowest theoretical performance cost at $27.27 TFLOPS ($10^{12}$ FLOPS), with a price of $3000 and a theoretical peak performance of 110 Tensor TFLOPS.

However, theoretical peak performance relies on the accelerating parts, and generally does not involve other hardware such as memory, network or I/O devices. An analysis of GPU performance, finds that theoretical predictions of maximum theoretical performance of three different GPUs to be higher by around 30% when compared to experimental results. Similar work on CPUs has found discrepancies between actual and theoretical maximum performance of generally within a factor of 10.

To take this into account, we shall here assume that theoretical cost figures (such as reported by hardware manufacturers are 10 times too low), setting current costs at $272.70 per TFLOPS. Getting to below $1 would therefore amount to a $\log(272.7)/\log(10) ≈ 2.4$ orders of magnitude reduction in cost per TFLOPS.

Interestingly, according to some estimates, a reduction in computing costs of roughly 2-3 orders of magnitude could place the cost of computation at the brain's communication performance in the range of human wages (around a couple hundred dollars per hour).

When will a TFLOPS cost less than $1?

This question resolves positively if a reputable source reports a TFLOPS for Tensor, or single/double precision performance to cost less than $1 using reliable performance measurement techniques. The usual theoretical price estimates given by manufacturers will be multiplied up by a factor of 10 for the sake of this question (hence requiring a $0.1 theoretical cost per TFLOPS for positive resolution).