Computing · Compute Affordability

Cost per GFLOPS

The invisible curve behind simulation, graphics, science, and AI.

• $1T+ → ~1¢ per GFLOPS
• Halving: 20mo → 45mo since 2011
• Driven by GPUs, now AI low-precision
• Not like-for-like: precision varies

The cost of arithmetic has collapsed so far that entire categories of research became software problems. From the room-sized ENIAC to commodity graphics cards, the price of a billion floating-point operations fell from over a trillion dollars to about a cent. GPUs accelerated the curve by turning parallel workloads into mass-market performance, and the AI boom has since made low-precision throughput the metric that matters. Every advance compounds the simulations, graphics, and models that cheap compute makes feasible. One caveat across eighty years: these points are not perfectly like-for-like. The oldest machines are measured in fixed-point or double-precision arithmetic and the newest in the lower-precision throughput GPUs now optimise for, so read this as an order-of-magnitude cost curve rather than a single benchmark held constant. The slope has also flattened: the price of a gigaflop halved about every 20 months from ENIAC through 2011, but has taken about 45 months per halving over the past decade. Note which way the caveat cuts — the recent points are measured in the low-precision arithmetic that flatters them most, and the curve is still slowing.

Source: Wikipedia: cost of computing (FLOPS)

Download the data: flops_per_dollar.csv · flops_per_dollar.json

All curves · The chart is drawn in your browser and needs JavaScript. Every number it plots is in the table above.