Work out your super computer power consumption in seconds: enter the wattage on your super computer's label, its hours of use per day and your electricity rate, then click Calculate. You get your daily, monthly and yearly cost and total kWh consumption straight away. Those same results also cover super computer electricity consumption. Also see tablet charger electricity consumption.
If you have ever wondered what it costs to keep the world's fastest machines running, super computer electricity consumption is the place to start: a modern exascale system draws tens of megawatts around the clock, roughly what a small city uses. In this guide you will see how much power a supercomputer draws, where that electricity goes inside the machine and the data center around it, and how to estimate the bill for a system of your own.
A supercomputer is a very large cluster of processors, memory and fast networking that works as one machine, and its power consumption is measured in megawatts rather than the watts you see on a laptop charger. A desktop computer might pull 100 to 500 watts under load. The largest exascale computers operating today are listed on the Top500 ranking at roughly 20 to 40 MW each, which is tens of thousands of times more than a desktop.
That range is easier to grasp next to everyday numbers. One megawatt can supply roughly a hundred typical US homes' worth of continuous demand once you account for the air conditioning, water heating and appliances that run through the day. A machine near 30 MW therefore competes with a small city for generating capacity, and the local utility has to plan a dedicated feed for it, just as it would for a steel mill or a large hospital.
Two different quantities get mixed up in almost every article on this topic, so keep them apart. Power is the rate at which electricity is drawn, in watts (W), kilowatts (kW) or megawatts (MW). Energy is power multiplied by time, in kilowatt-hours or MWh. A supercomputer drawing 8 MW for one hour uses 8 MWh. Your utility bills you for energy, while the engineers who design the building size the electrical feed for power.
| Unit | Equals | Typical meaning for a supercomputer |
|---|---|---|
| 1 watt (W) | 1 joule per second | A single low-power chip core |
| 1 kilowatt (kW) | 1,000 W | One server node with several accelerators |
| 1 megawatt (MW) | 1,000 kW | A few dozen racks of dense GPU nodes |
| 1 MWh | 1 MW sustained for 1 hour | The unit your electricity bill is built from |
| 1 GWh | 1,000 MWh | The scale of a large machine's yearly electricity |
The published power draw of the biggest machines gives you a feel for how the field has changed. The figures below are approximate values as reported on the Top500 list during the machines' benchmark runs; real draw moves up and down with the workload, so treat them as orders of magnitude rather than exact meter readings.
| System | Operator | Approximate power | Notes |
|---|---|---|---|
| Frontier | Oak Ridge National Laboratory | about 22 MW | First exascale machine, GPU-based |
| Aurora | Argonne National Laboratory | about 39 MW | Exascale, GPU-based |
| El Capitan | Lawrence Livermore National Laboratory | about 30 MW | Exascale, national security workloads |
| Fugaku | RIKEN, Japan | about 30 MW | CPU-only design, no GPUs |
Older machines tell the same story at a smaller scale. The Cray-1 of the 1970s was a landmark of its era and needed only a fraction of a megawatt. The Tianhe-2 in China and the K computer in Japan, both ranked first in their day, pushed the figure into the double-digit megawatts, and every new generation since then has asked for more generating capacity even as each watt does far more work.
Three forces push power requirements up. First, the chips themselves: a single modern accelerator is designed to dissipate 500 to 1,000 watts, and a large cluster contains thousands of them. Second, performance targets: the industry has moved from petaflops to exascale, or 1018 floating-point operations per second, so there are simply more transistors switching. Third, AI training has become a main customer for the biggest systems, and researchers who track frontier AI clusters report that their power capacity has been roughly doubling every year or so, a pace that outruns efficiency gains.
It helps to follow the electricity from the wall socket to the final byproduct. Nearly every watt a computer consumes ends up as heat, so a supercomputer is, in physical terms, an enormously expensive and very fast electric heater that happens to do arithmetic. The consumption splits into a handful of buckets. Also see tablet computer electricity consumption.
Cooling is the biggest non-computing line item, and good heat management is a design discipline of its own. Early machines relied on powerful air conditioning and raised floors; modern designs use liquid cooling, with water circulating through cold plates attached directly to the processors. A few builders go further with liquid immersion cooling, submerging boards in a dielectric oil, which removes the need for most cooling fans and cuts the electricity spent on them. Immersion has trade-offs: servicing a board means lifting it out of the fluid, and the fire and cleanup issues are not trivial.
Two ratios describe how much electricity is spent on everything besides the chips. Power usage effectiveness (PUE) compares the whole facility's draw with the draw of the IT equipment alone. A PUE of 1.0 would mean zero waste; real data center values range from about 1.1 for excellent liquid-cooled halls to 1.5 or higher for older air-cooled ones. The second ratio is overhead inside the machine, the share of draw taken by hosts, memory, fans and networking rather than the accelerator chips. Published figures for GPU server nodes suggest overhead multipliers somewhere between 1.3 and 2.5 on top of the accelerators' own rating.
You do not need a meter on the substation to produce a sensible estimate. Researchers who compile data on large AI clusters use a short chain of multiplications to turn a parts list into a power capacity figure. The same approach works for any accelerator-based machine. Related: how much electricity does a blender use.
$$P_{\text{facility}} = N_{\text{chips}} \times \text{TDP} \times k_{\text{overhead}} \times \text{PUE}$$
Here thermal design power (TDP) is the maximum power a chip is designed to draw, \(k_{\text{overhead}}\) is the system overhead multiplier, and PUE is the facility-level ratio described above. To go from power to yearly energy and cost, add the hours and the electricity price:
$$E_{\text{year}} = P_{\text{facility}} \times 8{,}760 \times u \qquad \text{Cost} = E_{\text{year}} \times \text{price per kWh}$$
The term \(u\) is average utilization, the fraction of peak draw the machine actually uses over the year. Because few clusters run flat out all day, annual electricity bill estimates that skip \(u\) tend to overshoot. As a rule of thumb, the Top500 machines measured during the Linpack benchmark have consumed around two thirds of their peak draw on average.
Suppose you are sizing a new training cluster with 6,144 accelerator chips, each rated at a TDP of 600 W. You assume a system overhead multiplier of 1.7, a PUE of 1.2, an average utilization of 65%, and an electricity price of $0.083 per kWh. Work through it in order:
| Quantity | Value | How it is obtained |
|---|---|---|
| Chip load | 3.69 MW | 6,144 chips × 600 W |
| IT load with overhead | 6.27 MW | 3.69 MW × 1.7 |
| Facility peak power | 7.52 MW | 6.27 MW × 1.2 PUE |
| Yearly energy (65% utilization) | 42.8 GWh | 7.52 MW × 8,760 h × 0.65 |
| Yearly cost at full draw, 100% | $5.47 million | 7.52 MW × 8,760 h × $0.083 per kWh |
| Yearly cost at 65% utilization | $3.55 million | 42.8 GWh × $0.083 per kWh |
The yearly energy of 42.8 GWh is about what 4,000 US homes use in the same period, assuming a household consumes roughly 10,500 kWh annually. If the same cluster were housed in an older hall with a PUE of 1.5, the facility peak power would rise to 9.4 MW, and at the same utilization and price you would pay roughly $0.89 million more per year. That single ratio is worth a lot of money.
Raw power consumption only tells half the story. A machine that draws twice the electricity but computes ten times faster is the better deal, which is why the field cares about energy efficiency, expressed as the performance-to-power ratio or gigaflops per watt. The Green500 list re-ranks the Top500 machines by exactly this measure, and its leaders are often not the fastest systems overall, so efficiency has become a separate race from speed.
$$\text{Efficiency} = \frac{\text{sustained performance (GFLOPS)}}{\text{power (W)}}$$
For the hypothetical 7.52 MW cluster above, imagine it sustains 250 petaflops on a benchmark. That is 250,000,000 gigaflops divided by 7,520,000 watts, or about 33 gigaflops per watt. A second cluster with the same speed but a PUE-inclusive draw of 12 MW would land near 21 gigaflops per watt, and its owner would pay roughly 60% more for electricity to get the same science done.
A major turning point was the adoption of GPUs for scientific computing. Graphics chips were designed to run many simple operations in parallel, which suits simulation and machine learning, and for the same amount of computing power they use far less energy than conventional CPUs. Today's exascale systems at the national laboratories depend on that shift; without it, the electricity needed for an exaflop machine would have been prohibitive.
After the arithmetic units, memory traffic is the next target. Moving a number from DRAM to a processor can cost more energy than computing with it, so vendors are bringing 3D stacked memory physically closer to the processor. Looking further ahead, optical interconnects are expected to replace copper wiring between racks, which should cut the energy spent on communication. Software helps too: schedulers that map the right task to the right processor at the right time keep hundreds of thousands of processors from wasting power waiting on each other.
The reason a supercomputer draws so much is its architecture. Instead of one very fast processor, supercomputing systems link hundreds of thousands of cores that work in parallel, each handling a slice of one huge problem. Since the 1990s that massively parallel layout has been the norm, and it means that energy usage scales with the number of nodes you add. Early examples such as the Blue Gene line showed that many modest, low-clock processors could beat a few fast ones on power efficiency, a lesson that later designs kept applying.
The first computers to carry the supercomputer label were vector machines, and a single one drew well under a megawatt. Massively parallel supercomputing replaced them because adding commodity processors was cheaper than designing a faster single one, but it multiplied the electricity bill: dozens of racks now draw what one vector machine once did, and coordinating that many parts costs power of its own. The software that divides the work, from operating systems to message-passing libraries, therefore has a direct effect on energy use.
Modern hardware and software are co-designed for efficiency. A Top500 entry is judged by its Linpack score, but a facility team also watches how many FLOPS each kilowatt buys. Specialised accelerators, task-aware schedulers and power-capping features let operators trade a few percent of speed for a bigger cut in power draw, which is often the right call once the electricity price is in the spreadsheet. In practice, that co-design is as much a part of energy efficient technology as the silicon itself.
The arrival of exascale computing changed the scale of the conversation. An exascale computer performs more than a quintillion calculations every second, and the three exascale supercomputers run by the national laboratories were each built around a power ceiling. Aurora, Frontier and El Capitan are the best-known examples, and earlier leaders such as the Summit machine and the Tianhe-1A showed the same pattern at smaller scale: every doubling of speed came with a hard limit on megawatts.
A newer category is AI supercomputers, clusters filled with accelerators and built chiefly to train large neural networks. Their sizing leans on the same formula given earlier, and operators compare them by average power over a training run as well as by peak. Because scientists and engineers now share these machines with commercial users, the demand for research time and electricity keeps growing together.
Supercomputer power usage is not abstract for the people nearby: a 30 MW exascale system needs its own dedicated substation. Local households can feel it through grid upgrades, cooling water use and an environment review before a new data hall opens. Utilities now ask developers to publish expected consumption early, so that the generating capacity and transmission lines are planned long before the first rack is powered on.
Scale makes more sense with a comparison. A gaming desktop under load pulls about 450 W, so one 7.52 MW cluster like our example equals roughly 16,700 such computers running flat out together. Yet that single supercomputer does what no pile of desktops can, because its fast interconnect lets every processor share results with every other one. Large-scale computing is valuable for that coordination, not for raw wattage.
For scientific research, the comparison is fair only when you divide by the work done. A parallel machine that finishes a climate simulation in hours rather than weeks may use less total energy per result than a smaller system running for far longer, because its shorter run time cancels some of its higher draw. This is why computing centres report energy per job alongside the headline megawatts, and why experts urge buyers to look at computing output per kilowatt-hour rather than a single power figure. When new hardware arrives, software teams re-tune their codes, because a program that ignores the new design can waste much of the saving.
For a facility manager, the energy requirements of a machine translate directly into operating costs. A common rule of thumb in the trade press has been that each continuously drawn megawatt costs on the order of a million dollars a year, though the real number depends entirely on local tariffs: your own price per kWh, contract terms and demand charges decide it. In the worked example above, 7.52 MW at $0.083 per kWh works out to about $0.73 million per megawatt-year at full draw, which is in the same neighbourhood but not identical.
Over a five-year service life the electricity can rival the purchase price of the hardware. That is a reason many sites choose locations with cheap hydroelectric power or cool climates, and why a new machine's contract often includes a performance-per-watt requirement rather than a pure speed target. Here are the main levers an operator can pull:
The same megawatts that stress a power grid also have a carbon footprint when they come from the wrong source. A machine that draws 7.52 MW for a year consumes 65.9 GWh at full draw; on a grid emitting 0.4 kg of CO2 per kWh, that is roughly 26,000 tonnes, an amount large enough to matter for any organisation with a climate target. Where the supply is coal-heavy, the footprint is far larger than on a grid dominated by hydro or nuclear power, which is why site selection and renewable contracts matter as much as chip choice.
Waste heat deserves attention as well. Because essentially all of the consumed electricity becomes heat, a handful of data centers now pipe it into district heating networks. That does not erase the emissions from generating the electricity, but it means the same energy serves two purposes. The wider debate on climate change and fossil fuels is part of the reason regulators and funders now ask for efficiency reporting alongside performance claims.
It is tempting to blame the headline machines, but they are few in number; the much larger total comes from billions of laptops, phones, game consoles and ordinary servers. The trade-off for a supercomputer is that each one draws megawatts but runs weather forecasting, climate research and artificial intelligence work, plus nuclear stockpile simulation and molecular research, that is hard to do any other way, so whether a job justifies its power budget is a policy question as much as a technical one.
Where does supercomputing go from here? Moore's law no longer delivers free efficiency gains each year, so improvements now come from specialised accelerators, denser packaging and smarter software. At the same time the appetite for compute is growing. Analyses of frontier AI systems using a log-linear regression report that their power requirements grew by around 1.9 times per year, with a wide confidence interval, which is a much faster trajectory than the typical efficiency improvement of a hardware generation.
The Department of Energy and its national laboratories have long tied exascale targets to an explicit power ceiling, aiming at machines in the 20 to 30 MW class rather than the hundreds of megawatts a naive scale-up would require. Expect the next wave to include more on-site generation, tighter coordination with the grid, and reporting of both peak and average figures. The distinction matters: the power capacity a utility must reserve is not the same as the electricity actually consumed over a year.
A fair comparison between two machines needs three numbers: speed, average draw and the facility's PUE. Quoting only one of them can make either machine look better than it is.
Overall, super computer electricity consumption comes down to chips, overhead and facility, and a supercomputer's appetite is tempered by how hard you run it. Use the formula and the worked example above to put numbers on any system you are evaluating, and always compare machines on performance per watt as well as on speed.
Looking across the field, supercomputing has become a question of energy as much as of speed. Every claim about FLOPS should come with a power figure, because computing capability that nobody can afford to switch on is of little use. The technology is advancing quickly, with better chips, smarter cooling and tighter software, and each advance in computing efficiency lets scientists and engineers run more research on the same electricity. Whether you manage a data hall, review a proposal or simply follow supercomputing news, the same habit applies: ask how many computers' worth of power it needs, what technology keeps it cool, and how much useful computing it delivers per kilowatt-hour.