jj4211

Worth noting that the fancy ice maker is just like a 30 dollar ice tray, with insulation and silicone mold.

It does take a long time to make a little ice compared to normal freezing.

Noe that the water can just be tap, and the equipment isn't super fancy. If the water looks clear, you can make clear ice from it.

It's called directional freezing, you stick water in freezer insulated on all sides so that it freezes from the top down instead of outside in.

If you have larger ice, you'll see the white stuff is in the middle, the last area to freeze. Directional freezing causes that to be at an end instead of in the center, and you either pull out the ice before the end freezes, ideally, or cut off the end.

I have an ice mold that doess this and it provides break off points to break the clear ice off the unclear ice.

It does take a while though and the bulky insulation takes a lot of room in the freezer.

For reference, a GB300 server is now at about 8.5kw for a single server. A fully populated NVL8 server is about 15kw. Looking online, looks like the ISS is about up to 90kw, so I guess I'm off and the actual number is something like 6 to 10 of these servers per ISS scale facility.

I would still argue this is a crazy overhead for what they would now consider meager capacity, with, luckily for nVidia, a pretty hard deadline where the very expensive equipment burns up without potential for extended lifetime use.

The solution to tackling it was to just not tackle it. See the various dangling cables where something was no longer needed and rather than removing the cable, screw it, it is lost to the entropy.

Man that madthumbs guy is really trying to make that a thing and it's kind of sad and lonely that he's off by himself pretending he has a community...

Oh, we are sharing workplace cable management, ok, here's a place I used to work at:

My favorite one was Musk gave.

So he said that the supply of a particular natural gas turbine part was constrained through 2030. Therefore, obviously, the simplest fix is to have thousands of starship launches with ISS-sized payloads and all the attendant crap.

For scale. roughly a two server datacenter needs to have solar and radiator about as big as the ISS.

Which is possible, but insane. However insane plain old datacenter is, just tons more insane.

Basically people see an address like fdec:46f7:9b7f:1::3:20 and run screaming away about the complexity, seeing the address as a comprehensive indication of complexity, even though the real challenges lie underneath.

The whole 'traditional ipv4 just has 0.0.0.0' stuck in front of it is essentially exactly the same idea as, say 64:ff9b::142.251.152.119. Now there's also the likes of ffff::142.251.152.119 but that's just so software can pretend to speak IPv6 when the OS is really doing only IPv4. So they needed another prefix to indicate the network doing the v6 to v4 translation instead of the OS.

Anyway, the thing is that while it cosmetically looks more similar, it's not really solving the fundamental compatibility situation. It just "looks nicer" because it sticks to dotted decimals. However in practice, would fdec:46f7:9b7f:1::3:20 really be somehow less usable than, say, 120.30.204.78.167.144.120.209? The simple reality is that the 4 octet decimal pushed human usability enough as it was, and going to sufficient octets just brings it out of mere mortal reach. If you did want to say have more friendly local network addresses (the vast vast majority of human memorized IP addresses), then technically you could have fd::1, fd::2, fd::3, and those would all work and be super easy to remember (the ULA RFC says you are supposed to toss in 40 bits of random for good reason, but if you were using 10.0.0.1 style addresses, you would be no worse off with fd::1, fd::2, etc). You can even trivially have them live alongside 'real' global IP addresses, but ignore them whenever you want to just hand type a local IP address. You can even have something like a hex DNS. fd::f00d, fd::beef, fd::d00d, and so many more for your pleasure.

There's more features in IPv6 but you can ignore them since they are mostly for the machines to wrangle (the fe80:: addresses for example).

Note while you have cosmetic similarities to ipv4 addresses, the actual challenging part of that is the packet format and various translations.

We actually have a number of existing schemes for ipv4 mapping onto larger address space and the attendant NAT requirements. The presentation of addresses in an ipv4 looking way is the least of the challenges.

So don't take IPv8 seriously, it is slop and even in theory it wouldn't add anything new except a different cosmetic look to raw addresses and shortening the address space for no good reason.

Don't think you grasp which side has the actual leverage here...

The US market is the one that has driven the truckloads of money that have resulted in the memory vendor stock increasing over 10-fold.

Even a pretty severe compromise or fine is totally worth it to keep the money hose going.

Working from memory here

Nowadays that's a pretty expensive way to work