Global computing capacity in terms of human brains

According to different estimates, the human brain performs the equivalent of between 3 x 10¹³ and 10²⁵ FLOPS. The median estimate we know of is 10¹⁸ FLOPS. According to that median estimate and our estimate of global computing hardware, if the world’s entire computing capacity could be directed at running minds around as efficient as those of humans, we would have the equivalent of 200-1500 extra human minds.⁷ That is, turning all of the world’s hardware into human-efficiency minds at present would increase the world’s population of minds by at most about 0.00002%. If we select the most favorable set of estimates for producing large numbers, turning all of the world’s computing hardware into minds as efficient as humans’ would produce around 50 million extra minds, increasing the world’s effective population by about 1%.⁸

Global computing capacity

2016-02-16 AI Timelines, Hardware and AI Timelines, Human Level Hardware 9

Computing capacity worldwide was probably around 2 x 10²⁰– 1.5 x 10²¹ FLOPS, at around the end of 2015.

Contents

Support

We are not aware of recent, plausible estimates for hardware capacity.

Vipul Naik estimated global hardware capacity in February 2014, based on Hilbert & Lopez’s estimates for 1986-2007. He calculated that if all computers ran at full capacity, they would perform 10-1000 zettaFLOPS, i.e. 10²²– 10²⁴ FLOPS.¹ We think these are substantial overestimates, because producing so much computing hardware would cost more than 10% of gross world product (GWP), which is implausibly high. The most cost-efficient computing hardware we are aware of today are GPUs, which still cost about $3/GFLOPS, or $1/GFLOPSyear if we assume hardware is used for around three years. This means maintaining hardware capable of 10²²– 10²⁴ FLOPS would cost at least $10¹³– $10¹⁵ per year. Yet gross world product (GWP) is only around $8 x 10¹³, so this would imply hardware spending constitutes more than 13% – 1300% of GWP. Even the lower end of this range seems implausible.²

One way to estimate global hardware capacity ourselves is based on annual hardware spending. This is slightly complicated because hardware lasts for several years. So to calculate how much hardware exists in 2016, we would ideally like to know how much was bought in every preceding year, and also how much of each annual hardware purchase has already been discarded. To simplify matters, we will instead assume that hardware lasts for around three years.

It appears that very roughly $300bn-$1,500bn was spent on hardware in 2015.³ We previously estimated that the cheapest available hardware (in April 2015) was around $3/GFLOPS. So if humanity spent $300bn-$1,500bn on hardware in 2015, and it was mostly the cheapest hardware, then the hardware we bought should perform around 10²⁰ – 5 x 10²⁰ FLOPS. If we multiply this by three to account for the previous two years’ hardware purchases still being around, we have about 3 x 10²⁰– 1.5 x 10²¹ FLOPS.

This estimate is rough, and could be improved in several ways. Most likely, more hardware is being bought each year than the previous year. So approximating last years’ hardware purchase to this years’ will yield too much hardware. In particular, the faster global hardware is growing, the closer the total is to whatever humanity bought this year (that is, counterintuitively, if you think hardware is growing faster, you should suppose that there is less of it by this particular method of estimation). Furthermore, perhaps a lot of hardware is not the cheapest for various reasons. This too suggests there is less hardware than we estimated.

On the other hand, hardware may often last for more than three years (we don’t have a strong basis for our assumption there). And our prices are from early 2015, so hardware is likely somewhat cheaper now (in early 2016). Our guess is that overall these considerations mean our estimate should be lower, but probably by less than a factor of four in total. This suggests 7.5 x 10¹⁹ – 1.5 x 10²¹ FLOPS of hardware.

However Hilbert & Lopez (2012) estimated that in 2007 the world’s computing capacity was around 2 x 10²⁰IPS (similar to FLOPS) already, after constructing a detailed inventory of technologies.⁴ Their estimate does not appear to conflict with data about the global economy at the time.⁵ Growth is unlikely to have been negative since 2007, though Hilbert & Lopez may have overestimated. So we revise our estimate to 2 x 10²⁰ – 1.5 x 10²¹ FLOPS for the end of 2015.

This still suggests that in the last nine years, the world’s hardware has grown by a factor of 1-7.5, implying a growth rate of 0%-25%. Even 25% would be quite low compared to growth rates between 1986 and 2007 according to Hilbert & Lopez (2012), which were 61% for general purpose computing and 86% for the set of ASICs they studied (which in 2007 accounted for 32 times as much computing as general purpose computers).⁶ However if we are to distrust estimates which imply hardware is a large fraction of GWP, then we must expect hardware growth has slowed substantially in recent years. For comparison, our estimates are around 2-15% of Naik’s lower bound, and suggest that hardware constitutes around 0.3%-1.9% of GWP.

Such large changes in the long run growth rate are surprising to us, and—if they are real—we are unsure what produced them. One possibility is that hardware prices have stopped falling so fast (i.e. Moore’s Law is ending for the price of computation). Another is that spending on hardware decreased for some reason, for instance because people stopped enjoying large returns from additional hardware. We think this question deserves further research.

Implications