Performance per watt, 2000 to today

The 382 systems in this dataset that report both a measured Rmax and a power draw delivered a median of 0.20 GFlop/s per watt when they first ran in 2005 to 2009, and 42 in 2025 to date: 206 times as much. A straight line through the logarithm of the points doubles every 2.3 years (R² 0.81). The most efficient machine here, Otus, delivers 68. Accelerated systems lead by 3.8 times in the latest era.

This is the ranked world only. Efficiency here is a measured FP64 benchmark result divided by the power drawn during that run, so it exists only for systems that submitted one: 382 of the 735 systems with an Rmax in this dataset. The AI clusters being built today mostly report low-precision peak figures and no power draw at all, and are not on this page, which is why the line it draws is the efficiency of the machines a ranking can see, not of compute in general.

Phase 1 seed dataset, compiled by hand. These rows were built from public operator, laboratory and vendor sources. A mechanical second-reader pass has since fetched every cited source: 1092 of 1153 systems have a readable citation that names them, and 40 are genuinely weakly sourced. Every claim carries its source and a confidence tier. Treat anything below verified as a lead, not a citation.

CPU-only
Accelerated

Each dot is one system: GFlop/s per watt on a logarithmic axis, so equal vertical steps are equal multiples. The line joins the median of each five-year group with at least five systems.

Download this chart's data (CSV) ↓

Table view: Measured performance per watt by first-light year

Years	Systems	Median GFlop/s per W	Accelerated (n)	CPU-only (n)	Most efficient
2000 to 2004	1	0.33	-	0.33 (1)	BlueGene/L 0.33
2005 to 2009	29	0.20	0.32 (3)	0.20 (26)	Hexagon 0.47
2010 to 2014	80	0.91	1.5 (28)	0.87 (52)	L-CSC 5.5
2015 to 2019	124	3.0	4.1 (55)	2.5 (69)	A64FX Prototype 17
2020 to 2024	115	22	37 (77)	5.5 (38)	JEDI 67
2025 to date	33	42	47 (26)	12 (7)	Otus 68

How this was calculated

Efficiency is Rmax in PFlop/s times 1,000, divided by power in kW (one PFlop/s at one thousand kilowatts is exactly one GFlop/s per watt). A system is included if it has both figures and a first-operational date, and is placed at that year. Power is the figure recorded on the system's row, almost always the measured draw reported with a benchmark run, not a facility's connected capacity.

Two kinds of row are left out because their Rmax and power describe a later, larger configuration than the year they would be plotted at: anything recorded as upgraded, and Pleiades, whose 2010s figures would otherwise appear as an efficiency jump in 2008. A system expanded in place and not marked upgraded is still dated to its first light, so a few early points may sit slightly too early; the five-year medians are robust to a handful of them.

“Accelerated” means the system has at least one accelerator edge in the component graph. The doubling time is the least-squares slope of log base two of the efficiency on the year, so it is a trend through scattered points, not the pace of any one supplier.

The efficiency frontier

The most efficient system to have run as of each date, so each row beat every system before it.

First light	System	Country	GFlop/s per W	Measured Rmax	Power
2004	BlueGene/L	USA	0.33	478 TF	1.4 MW
2007	JUGENE	Germany	0.38	191 TF	504 kW
2008	Hexagon	Norway	0.47	160 TF	342 kW
2010	Tianhe-1A	China	0.64	2.6 PF	4.0 MW
2011	HokieSpeed	USA	0.95	120 TF	126 kW
2012	SANAM	Saudi Arabia	3.0	533 TF	179 kW
2014	L-CSC	Germany	5.5	317 TF	57 kW
2016	Sakura (PEZY)	Japan	17	825 TF	50 kW
2017	Suiren2	Japan	17	798 TF	47 kW
2019	A64FX Prototype	Japan	17	2.0 PF	118 kW
2020	MN-3	Japan	41	2.2 PF	53 kW
2021	Frontier TDS	USA	62	19.2 PF	309 kW
2022	Henri	USA	65	2.9 PF	44 kW
2024	JEDI	Germany	67	4.5 PF	67 kW
2025	Otus	Germany	68	4.7 PF	68 kW

The read

The curve is steeper than the headline suggests, and not smooth. Efficiency crept through the 2000s, when the median system delivered about a fifth of a GFlop/s per watt, then climbed by a factor of 15 between the 2005 and 2015 groups. The step that shows in the frontier is accelerators: the first GPU-heavy machines (Tianhe-1A, then SANAM and L-CSC) opened a gap over CPU-only designs that has not closed, and every machine to set the frontier since 2020 is accelerated.

The gap between the accelerated and CPU-only medians is the practical cost of choosing a design that does not use them: in the latest group the CPU-only median is 12 against 47. Some of those CPU-only systems are deliberate, bought for codes that do not map onto GPUs, so the lesson is not that they are wrong but that they pay for it in power.

What the page cannot say is whether the large AI campuses are efficient. They are measured in megawatts of facility capacity and in low-precision petaFLOPS, neither of which belongs on this axis. The build pipeline is where their power appears; no equivalent measured efficiency exists for them yet.