Rambus, Microsoft Put DRAM into Deep Freeze to Boost Performance
nextplatform.com
nextplatform.com
Cryogenic computing wasn't a bust because the technology didn't work. It does. It was a bust because silicon really hates to transition between cryogenic temperatures and room temperature. If you transition it quickly (say you pull a card out of the liquid nitrogen and start working on it) it will crack as it warms up unevenly. As a result you needed anywhere from 20 to 48 hours to get a card from 'cryo' temp to 'room' temp, and while you could cool a bit faster it still took longer than just dumping it in LN2. So repairs and maintenance were multi-day affairs. Compare that to a modern AWS, Google, or Azure data center where a system fails, a tech can skate out to it with a new motherboard, pull the old one, put the new one in, and poof you're back on line in under 30 minutes.
As a result cryo computers either had to have failure rates that were so low that a repair that required transitioning the hardware through a cold/warm/cold cycle rarely happened, or you had to have enough extra hardware to support your base load while part of it was slowly going through the cold/warm/cold cycle.
I don't know how much IBM spent, but it was a lot and they never cracked that nut.
Modularity is better now than it was then. I could very much see the possibility of having pressure-sealed internally-cryo-maintained enclosures (like the pressure-sealed internally-helium WD hard drives), as hot-swappable modules with e.g. Thunderbolt 3.1 connectivity to connect the enclosure-module to your other components. This approach would work fine for, say, enclosing NVMe. And, conveniently, repair would be facilitated by the inside of the enclosure staying cold even after you yank it out of the machine. A daughterboard in a thermos!
For actual DRAM DIMMs, though, you'd have to "enclose" the motherboard itself into such a module—at which point that "module" is the computer, and you're more just talking about the old blade-server paradigm of externalizing as many components as possible to live on the rack, so that the server blades could be treated as commodity parts and swapped out like any other.
You can make a unit operate in cryogenic temperatures and, when it fails, power it down and manage its cooling long enough it can be removed safely. When repaired, just re-add it and cool it down until you can safely bring it online.
Ultimately its just dollars per compute and it seems highly dubious to me that supercooling your components, even if you accept high failure rates and long lead times, is going to move that needle in a positive direction.
But isn't "the temperature card" something you only get to play once as well? Once you've adopted supercooled computing, where do you go from there?
(This guy knows there's such a thing as absolute zero, and that you can't just keep getting colder, right?)
Sorry, I guess I still have antipathy towards Rambus since the 1990s. =)
https://en.wikipedia.org/wiki/Rambus#Patent_Lawsuits
I just remembered the Intel/Rambus boards that all had to be recalled because of Rambus issues (IIRC, the issue wasn't with the ram, but actually the boards themselves. Many of the boards were given to employees who took the ram and chucked the boards).