3 ms·

>This won't replace metal layers in CPU for a really long time. Superconductors are hard enough to make, CPUs are absurdly hard to make, and the wins on power s

>This won't replace metal layers in CPU for a really long time. Superconductors are hard enough to make, CPUs are absurdly hard to make, and the wins on power savings aren't very large. If people make superconducting chips, it will be ones where the superconductors do active switching, what is much farther away and can enable much faster CPUs too. I would've thought the main wins would've been reduced heat generation. Like you said, power savings would be negligible. But at the data center scale I'd imagine that reduced heat would result in reduced air conditioning power costs. And potentially with thermal constraints removed, it would allow for more compact packaging per server as well. Like you said though, I doubt it would happen any time soon.

cogman10 3y ago
> I would've thought the main wins would've been reduced heat generation. Most of the heat in a data center is coming from power burned in the devices themselves not the power transmission to those devices. Sort of like how a space heater is hot at the heater portion, not the power cable going into the heater. You aren't losing more than a W or so to transmission for every kW of power delivered. (in fact, you are generating more heat from the AC->DC transformation) Typically, power cords and wiring is 15 AWG, which has a 10 milli-ohm/meter resistance and runs as 120V AC (maximum of 15->20A). So, 1000m of power cord running at full load would result in 150->200W of heat from the power cable. Meanwhile the server is generating 1.6->2.2kW of heat. (Assuming a single very long 1kM route is servicing the server). Cut the cabling distance to a couple of meters and you can see why nobody worries at power consumption at that point.
- somerandomqaguy 3y ago
  I meant in the integrated circuits rather then power supply.
marcosdumay 3y ago
You can start to get to the answer if you look why metal layers aren't made with silver. If metal resistivity was that important, nobody would pass the chance of almost doubling the conductivity with a much smaller change than people are proposing here. The reason they don't is because yeild is the one most important variable on the entire process. And silver adds enough complexity to decrease it. (A few chips do have silver layers. People manage to use them when the process is mature enough and the added complexity gets tamed. Including something like YBCO into a high-performance chips manufacturing process is a half-a-century project; LK-99 can't even be reliably done yet.)

cogman10 3y ago

> I would've thought the main wins would've been reduced heat generation. Most of the heat in a data center is coming from power burned in the devices themselves not the power transmission to those devices. Sort of like how a space heater is hot at the heater portion, not the power cable going into the heater. You aren't losing more than a W or so to transmission for every kW of power delivered. (in fact, you are generating more heat from the AC->DC transformation) Typically, power cords and wiring is 15 AWG, which has a 10 milli-ohm/meter resistance and runs as 120V AC (maximum of 15->20A). So, 1000m of power cord running at full load would result in 150->200W of heat from the power cable. Meanwhile the server is generating 1.6->2.2kW of heat. (Assuming a single very long 1kM route is servicing the server). Cut the cabling distance to a couple of meters and you can see why nobody worries at power consumption at that point.

somerandomqaguy 3y ago
I meant in the integrated circuits rather then power supply.

marcosdumay 3y ago

You can start to get to the answer if you look why metal layers aren't made with silver. If metal resistivity was that important, nobody would pass the chance of almost doubling the conductivity with a much smaller change than people are proposing here. The reason they don't is because yeild is the one most important variable on the entire process. And silver adds enough complexity to decrease it. (A few chips do have silver layers. People manage to use them when the process is mature enough and the added complexity gets tamed. Including something like YBCO into a high-performance chips manufacturing process is a half-a-century project; LK-99 can't even be reliably done yet.)