How GPU Price-Performance Doubles Every 2.5 Years
Nobody buys a graphics card hoping it'll become obsolete next year. We buy because we expect improvement — and lately, the data gives us good reason to.
Using a dataset of 470 graphics cards released between 2006 and 2021, the amount of floating-point operations per second per dollar — FLOP/s per $ — doubles every ~2.5 years. That's the headline finding from Epoch AI's 2022 report, and it's the kind of number that gets quoted in AI timelines, hardware forecasts, and venture capital decks. But the headline masks a split: models typically used in ML research see a faster 2.07-year doubling time, while top-performing GPUs at any given moment improve more slowly, at 2.95 years.
The dataset itself is a merge. One piece came from the Median Group (223 Nvidia and AMD GPUs), another from Sun et al. (2019) (413 GPUs). Duplicates were stripped, and anything before 2006 was dropped, not because the data was unavailable, but because it felt incomparable. Pre-2006 GPUs often lacked reliable FLOP/s figures, and the general-purpose GPU era essentially begins with CUDA in 2007. The final set contains 470 GPUs from AMD, Intel, and Nvidia.
The regression through all 470 models yields a doubling time of 2.46 years (95% CI: 2.24 to 2.72 years). For ML GPUs, that subset of 26 models commonly used to train machine learning systems, the doubling time shortens to 2.07 years (95% CI: 1.54 to 3.13 years). Top-performing GPUs, selected as the highest FLOP/s per dollar each month, show 2.95 years (95% CI: 2.54 to 3.52 years). Those confidence intervals are wide enough that the differences aren't statistically significant, but the directional split is clear.
How does this compare to the laws named after Moore, Huang, and Cotra? Moore's law, the 2-year transistor-density doubling, translates to a 2-year FLOP/s doubling, so GPU price-performance has generally been slightly slower. Huang's law, which claims a 25x improvement every 5 years (roughly 1.08-year doubling), is strongly overstated when price-per-dollar is the metric. Cotra's 2020 prediction of a 2.5-year doubling sits almost exactly where this data lands. Prior work by Bergal (2019) found a 4.4-year doubling in FP32, and the Median Group's earlier estimate was 1.5 years, both well outside this dataset's interval.
Precision matters. The report had FP16 data on 91 GPUs, finding a 2.30-year doubling time (95% CI: 1.69 to 3.62 years). That's not significantly different from the FP32 trend, which means the price-performance improvement is roughly the same whether you're using half-precision or full precision. That stands in contrast to Bergal's 1.8-year FP16 FMA doubling, but the Epoch authors chose FP32 as their primary metric because the two trends appear comparable.
The ML GPU subset improves faster than the general pool, but the authors caution that this could reflect labs spending more on top GPUs over time rather than hardware improving faster across the board. If you're buying the best GPU available each month, you'll see a 2.95-year doubling, statistically significantly longer than the typical GPU. That distinction matters for anyone trying to project compute availability for future training runs.
The authors settle on ~2.5 years as the reliable estimate. Below 2 years feels aggressive; above 3 years feels at odds with the data. A doubling time of 2.5 years, the same slope used in Cotra 2020, feels like the best single number to carry forward.
What's striking is how compact the story is. Forty-seven hundred GPUs, fifteen years of retail and OEM data, and the central trend resolves to a single interval: 2.5 years. It's the kind of number that feels inevitable once you see it, but which you'd have to do the work to discover. And for a field that likes to project exponential growth, having a grounded historical anchor seems worth keeping around.
This report does not attempt to project future GPU price-performance. It takes stock of the recent historical trend. The code that generated all the figures is publicly available, even if the data itself isn't published. For anyone building timelines, estimating training run costs, or just wondering why GPUs keep getting better for the same money, the 2.5-year doubling is the number that the data actually supports.
Source
Hobbhahn, M. and Besiroglu, T. (2022). "Trends in GPU price-performance." Published online at epoch.ai. Retrieved from https://epoch.ai/publications/trends-in-gpu-price-performance [online resource]. Accessed 29 Aug 2026.
