Doing all the computer’s calculations twice—first forward and then backward—can drastically reduce the problem of AI overconsumption of power.
It all goes back to a decision made decades ago about the deep workings of computer logic and about the how these machines erase dataa process that inevitably produces a large amount of residual heatstarts by explaining .
For a long time, we stuck with computers that waste energy.
But with the rise of artificial intelligence, which has raised computing power demands to new heights, this seemingly inconsequential decision could be about to go wrong.
We may even have to redesign computing from the ground up. But there is a solution to end this problem: make the processors do everything twice, one for the front and one for the back.
“A reversible computing can be much more energy efficient than conventional computing and is potentially the way we should have built computers”, he guarantees Hannah Earley, from British reversible computing company Vaire Computing.
The technique will increase energy efficiency, and in fact has been around since the 70sbut was never put into practice due to constant improvements in traditional computing.
The energy consumption of AI technologies used every day is truly alarming: the daily energy consumption of ChatGPT, for example, it is equivalent to that of about 180,000 homes in the USand AI’s energy needs are increasing daily, putting a strain on the power grid.
Data centers and cloud computing, which are used for various services including AI, are already responsible for aroundand 2% of global emissions of greenhouse gases, a higher percentage than in the entire aviation sector.
And computers spend much more energy producing heat than performing calculations. The key intervention is then change the way we use logical operations that underlie computer programs.
Reversible computers take advantage of thermodynamic peculiarities in another way: move the electrons, which is the physical process underlying everything a computer does, more slowly than traditional computers.
This means that, viewed as a physical system, a reversible computer is never far from equilibrium, with fewer electrons being tossed back and forth.
Like this, less likely to use a lot of energy to return to its basic state of equilibrium, which again reduces waste.
“We know of no fixed limit to the amount of energy savings that can be achieved with reversible computing. Ten times, a hundred times, a thousand times — in theory, it’s always like this,” he says Michael Frankalso from Vaire Computing.
However, Jacques Carette, of McMaster University in Canada says that a completely logically reversible computing language would appear “extremely strange” to a software programmer with conventional training and that it would be a challenge to build a language of this type from scratch.
Vaire Computing is trying to reverse (literally) that. “The goal we are trying to achieve is make it look as normal as possible.”
Earley says Vaire last year patented a new type of resonator which will make its chips much more practical than any of the prototypes from the 1990s.
“The tools researchers used were more primitive at that time and designing the chips was a more manual and labor-intensive process,” says Frank. “To the modern tools can save us timebut some of these tools also need to be rewritten for reversible design. That’s why, at Vaire, we are build the tools we need as we move forward.”
“It’s great to see attempts to develop new technologies to address computing power consumption, but the challenge will be speed with which they can be placed on the market, because the demand… is happening now”, says David Mytton from the University of Oxford.
“In five years, there could be much more diversity, a Cambrian explosion of niche ideas”, they say Patrick Coles e Gavin Crooks, from Normal Computing, agrees: “There will be an avalanche of special chips. Instead of competing, they will help each other.”
