5 ms·
And it will go from 9% to 90% over the century. Big tech companies might as well avoid the middleman and evolve into the power generation business, although it'
by OutOfHere 2y ago
And it will go from 9% to 90% over the century. Big tech companies might as well avoid the middleman and evolve into the power generation business, although it's not hard with an open design of a solar farm in the desert.
- goda90 2y agoUnless we come up with newer more efficient algorithms and specialized hardware for implementing AI.
- talldayo 2y agoglances at Bitcoin mining ASICs Any day, now.
- jncfhnb 2y agoBitcoin is definitionally inefficient to mine. That’s the whole point.
- talldayo 2y agoThe point is that difficulty is adjusted in accordance with network saturation to prevent a 51% attack. The network is not inherently inefficient unless you're attempting to monopolize it - you are still rewarded for using more efficient mining methods than your competitors. Which leads us to the GPUs. For a while everyone said ASICs (and the courageous ones, FPGAs) were the future of mining because they were inherently more efficient than GPU hardware could ever be. But it turns out that it's the other way around - mass-manufactured GPUs have access to more efficient silicon than any FPGA on the market could hope for. Similarly, AI chips that want to beat Nvidia on efficiency terms first have to climb a mountain of power usage that is inherent to their exponentially less-dense hardware.
- yitternet 2y agoI think it dpends on the PoW system used. On a basic level you have memory hardening, and that can be used to force usage of bulk memory on GPUs or bulk storage like with Chia. For the more complex PoW you can use speculative execution to make CPUs most performant like RandomX. With bitcoin its more arbitrary but you can see how they could hard-fork to a different hardware-favoring system if desired.
- talldayo 2y agoWith Bitcoin specifically, I literally cannot imagine a serious number of people considering a hard-fork that removes the sole reason it's attractive as a currency.
- rcxdude 2y agoThe observation is that even if you get better at mining, all that happens is that the miners start to put in the same amount of effort but using the new technique: calculating hashes with less power or faster doesn't actually improve the resource usage of bitcoin, because the resource usage is tied to the rewards given to miners, not to the actual act of mining. (Importantly, it's not tied to the value of an attack! Nothing about bitcoin's now-ossified design actually ensures any particular level of resistance to attack, other than at the moment it's "very high" because there's still quite expensive coins being mined and people buying them) And the CPU/GPU/FPGA/ASIC thing is something that depends on the proof-of-work. Bitcoin went for the naive solution and it meant ASICs were the optimal solution once a certain amount of income was available. Later cryptocurrencies learned from this and specifically designed their proof of work systems to be optimal on a GPU, or even a CPU, because making ASICs the optimal solution is generally pretty bad if you want a diversity of miners.
- dylan604 2y agoGiant buildings in the desert with solar cells on top. No need for distribution lines. How many people are actually necessary to staff one once it is up and running? Bury the building with the desert soil for insulation.
- qeternity 2y agoI think you massively overestimate the generation capacity of a PV cell. If you cover all the surface area of a datacenter with PV cells, it will generate a small fraction of the power required.
- dylan604 2y agoI didn't illustrate the thought clearly. I didn't mean that the building's roof would be the only source of PV cells. The solar farm could grow all around the data center, but the roof would not need be a hole in the farm.
- OutOfHere 2y agoYour idea could perhaps work only in space, a bit closer to the sun, operating at peak power generation 24x7. On the ground it takes a lot more area in panels than is used by a datacenter.
- dylan604 2y agoOkay, so have AI build a nuclear plant in the desert so when they run it at 115% on the reactor, it's in the desert away from people. It's not like they'd be concerned about downwinders. Put the plant and the data center right next to each other again to avoid distribution lines. If they want power, make them figure it out. It should not EVER mean that normies have to sacrifice because AI wants
- OutOfHere 2y agoYou are all over the place. In any case, nuclear fission energy doesn't work without a river providing steady cold water for cooling. There is no such water in the desert.
- llm_trw 2y agoSolar power is the least suited power source for dataceners other than wind. You need nuclear reactors or other base line power generation energy sources.
- bo1024 2y agoFor inference, I agree. For training - the much larger use of electricity - I think solar might be great.
- llm_trw 2y agoTurning off the data center for 2/3rds of the year when the sun isn't shining sure is an idea.
- OutOfHere 2y agoIn case you forgot, the sun is always shining on the planet, 24 hours a day, every day. The location varies. A global transmission network will fix the availability issue.
- llm_trw 2y agoAnd nuclear power plants would solve it for a fraction of the cost without the chance of a world wide power grid failure.
- dalyons 2y agoHow so? Nuclear power is the most expensive mainstream form of power by a significant margin.
- llm_trw 2y agoThe price of solar power at midnight is infinite. By comparison nuclear is dirt cheap. People understand the fact it costs money to dig out coal from the ground and bring it to a power station. Yet they don't understand it costs money to take power generated at noon and shift it around. This is never included in the price for solar or wind.