Comments on: Getting To Zettascale Without Needing Multiple Nuclear Power Plants

By: emerth

emerth — Tue, 14 Mar 2023 14:34:56 +0000

So the supercomputers that will be sble to solve climate change will dissipate so much heat that they will fry the planet.

I sincerely hope the first one built is named “Malthus”!

By: HuMo

HuMo — Mon, 13 Mar 2023 19:39:41 +0000

In reply to Stephen Elbert.

… twistronics notwithstanding (magic-angle graphene …), or, hopefully, other superconducting switches with humorous, yet unanimously inspirational, names!

By: Tim Randles

Tim Randles — Thu, 09 Mar 2023 03:26:52 +0000

Chasing FLOPs is a stupid stunt. Go to the Top500 list and calculate efficiency. We already have way more FLOPs than we can use when the most computationally-dense benchmark gives us 60-65% efficiencies and something that looks like a real application (kind of) is at about 1%. Machines have been unbalanced for decades and FLOPs is a MEANINGLESS number.

By: Stephen Elbert

Stephen Elbert — Wed, 08 Mar 2023 19:42:59 +0000

In reply to HuMo. A problem with superconductor circuits is there are no efficient diodes.

By: HuMo

HuMo — Sat, 04 Mar 2023 15:17:52 +0000

Great summary of Su’s excellent ISSCC 2023 plenary talk (that kids can also watch on youtube, if so inclined)! The challenges we’re facing with energy dissipation come (I think) from our insufficient comprehension of the physics of matter (as a species). It would help, for example, if we could figure out room-temperature superconductivity, in a usable way. It would also be great if we could produce controllable electrical currents directly from nuclear processes (fusion, fission, or something other), rather than generating heat from them, using that heat to produce steam, using the steam to run a turbine, and then using the turbine to drive an electrical generator (alternator). Hopefully we can use Exascale (and beyond) computational tools to further our fundamental understanding of matter, and help us get to Zettascale (and beyond), energywise.

By: JayN

JayN — Sat, 04 Mar 2023 04:18:09 +0000

co-packaged optics, 3d fabrication, standardized chip to chip interfaces, domain specific accelerators, silicon interposers … looks like Intel’s research projects from the last 15 years.

I question whether Zettascale is possible by the end of the decade if relying primarily on hardware advancements. I think Intel’s (R. Koduri’s) plan relies primarily on software advances and I vaguely recall Dr Su putting more emphasis on hardware/software co-optimization in her presentation.