@mnovendstern I suspect that solar might beat nuclear for AI ... e.g. 25% of the land area of Arizona gives you 6000GW of power using 30% efficient solar, obviously that's an average figure so double it for the daytime.
Problem is the materials to make that much solar is a challenge.
Problem is the materials to make that much solar is a challenge.
@mnovendstern AI can be flexible in location and timing which is good for renewables.
@mnovendstern Power consumption beyond 5000GW seems like it will require nuclear... but at that point cooling becomes a serious issue.
@mnovendstern Even 1000GW might be a challenge TBH. But I think solar can do that, it's just a materials question.
1000GW is 10ยนยฒW or 10ยนยน human brain equivalents.
To get to a supercomputer equal to a Kardashev-I civilization (10ยนโทW) you have serious problems with waste heat
1000GW is 10ยนยฒW or 10ยนยน human brain equivalents.
To get to a supercomputer equal to a Kardashev-I civilization (10ยนโทW) you have serious problems with waste heat
@mnovendstern 1000GW is roughly equal to the total global cumulative installed capacity... but this is increasing incredibly fast. So maybe by 2030 we'll have 10,000GW of global installed capacity and a cost of $0.2/W so a 1000GW solar system would be $200bn just for the panels
@mnovendstern However it seems that the capital cost of compute vastly exceeds the cost of those solar panels. An A100 consumes 250W (i.e. about $250 in solar panels + storage/cooling/overheads) but it sells for a whopping $30,000 - 100x the cost of its power.
Loading suggestions...