That’s a good start. Let’s see; if we start with platters holding 0.6 terabytes in 2014, and assume an annual 15% increase, then platters start hitting the petabyte range in… 2070ish? Does that look about right?
(Yes, I know any particular percentage can be argued against. This is for fiction—I’m going for reasonable plausibility, not for betting on prediction-market futures.)
1.15^50 = 1084, so given the 15% rate of growth you’ll have an increase of about three orders of magnitude in fifty years.
In this specific case, though, the issue is whether rotating-platter technology will survive. In a way it’s a relic—this is a mechanical device with physical objects moving inside, at pretty high speed and with pretty tiny tolerances, too. Solid-state “hard drives” are smaller, faster, less power-hungry, and more robust already. Their only problem is that SSDs are more expensive per GB, but that’s a fixable problem.
True—but for my purposes, having /some/ number, even if it’s known to use poor assumptions, is better than none. I’m looking for things like “in which decade does a program requiring X MIPS become cheaper than minimum wage?” and “when can 100 petabytes be stuffed into ~1500 cm^3 or less, and how much will it cost?”. Which crossovers happen in which order is more interesting than nailing down an exact year.
when can 100 petabytes be stuffed into ~1500 cm^3 or less
Well… a 200Gb microSD card already exists. So you need five of them per 1Tb, 5000 per 1Pb and 500,000 per 100 Pbs.
A microSD card is 11 x 15 x 1 mm = 165 mm3 = 0.165 cm3 and some of that is packaging and connectors.
500,000 x 0.165 = 82,500 cm3. You wanted 1,500? That’s only about 50 times difference and getting rid of all that packaging and connectors should get you to about 30 times difference, more or less.
So the current flash memory density has to improve only by a factor of 30 or so to get you to your goal. That doesn’t seem to be too far off.
The fun task of calculating the bandwidth of one of those stuffed to the gills with contemporary microSD cards is left as an exercise for the reader :-)
Depends on the use case, I guess. The memory is non-volatile and the start-up time is negligible. If you only access one petabyte of memory within some time period, the other 99 can stay switched off and emit no heat.
I’m looking for things like “in which decade does a program requiring X MIPS become cheaper than minimum wage?”
In about a decade we will have machines that cost less than $10,000 and can run roughly brain sized ANNs. However, this prediction relies more on software simulation improvement rather than hardware.
“when can 100 petabytes be stuffed into ~1500 cm^3 or less, and how much will it cost?”.
Storage is much less of an issue for brain sims because synaptic connections are extremely compressible using a variety of techniques. Indeed current ANNs already take advantage of this to a degree. Also, using typical combinations of model and data parallelism a population of AIs can share most of their synaptic connections.
That’s a good start. Let’s see; if we start with platters holding 0.6 terabytes in 2014, and assume an annual 15% increase, then platters start hitting the petabyte range in… 2070ish? Does that look about right?
(Yes, I know any particular percentage can be argued against. This is for fiction—I’m going for reasonable plausibility, not for betting on prediction-market futures.)
1.15^50 = 1084, so given the 15% rate of growth you’ll have an increase of about three orders of magnitude in fifty years.
In this specific case, though, the issue is whether rotating-platter technology will survive. In a way it’s a relic—this is a mechanical device with physical objects moving inside, at pretty high speed and with pretty tiny tolerances, too. Solid-state “hard drives” are smaller, faster, less power-hungry, and more robust already. Their only problem is that SSDs are more expensive per GB, but that’s a fixable problem.
True—but for my purposes, having /some/ number, even if it’s known to use poor assumptions, is better than none. I’m looking for things like “in which decade does a program requiring X MIPS become cheaper than minimum wage?” and “when can 100 petabytes be stuffed into ~1500 cm^3 or less, and how much will it cost?”. Which crossovers happen in which order is more interesting than nailing down an exact year.
Well… a 200Gb microSD card already exists. So you need five of them per 1Tb, 5000 per 1Pb and 500,000 per 100 Pbs.
A microSD card is 11 x 15 x 1 mm = 165 mm3 = 0.165 cm3 and some of that is packaging and connectors.
500,000 x 0.165 = 82,500 cm3. You wanted 1,500? That’s only about 50 times difference and getting rid of all that packaging and connectors should get you to about 30 times difference, more or less.
So the current flash memory density has to improve only by a factor of 30 or so to get you to your goal. That doesn’t seem to be too far off.
The fun task of calculating the bandwidth of one of those stuffed to the gills with contemporary microSD cards is left as an exercise for the reader :-)
Don’t forget about the sheer amount of waste heat used by such an array were it actually on.
Depends on the use case, I guess. The memory is non-volatile and the start-up time is negligible. If you only access one petabyte of memory within some time period, the other 99 can stay switched off and emit no heat.
In about a decade we will have machines that cost less than $10,000 and can run roughly brain sized ANNs. However, this prediction relies more on software simulation improvement rather than hardware.
Storage is much less of an issue for brain sims because synaptic connections are extremely compressible using a variety of techniques. Indeed current ANNs already take advantage of this to a degree. Also, using typical combinations of model and data parallelism a population of AIs can share most of their synaptic connections.