I am the specific target audience for these drives.

And ... they are a TERRIBLE idea.

Assume PCIex4 v5.0 for the interface. That's a theoretical 15.75GB/sec. To read this drive sequentially would take 4.25 hours.

This is so slow it's absolutely useless for AI. Assume for a moment I loaded 8 of these into a 1u chassis. 800Gb XDR InfiniBand would be too slow, a double link would work. But you would be better off building half-U trays with four drives and an 800Gb link.

That said, let's say you had half a rack of that. That would be 48x245TB or about 12PB. And remember this is performance storage, not reliable storage. Everything here should be treated as entirely volatile... it's just cheap/slow RAM, it's not bad.

I think overall, I would architect a similar system on 64TB sleds because with the exception of rack space and power (and the drives use no power next to GPUs), 64TB drives destroy 245TB drives in every way.

Once we hit PCIe v9.0 or so and 4Tb Ethernet or InfiniBand, then 245TB will start making sense.

If Micron wanted a serious product, they would have dropped U.2 in favor of Ethernet or InfiniBand.

> Throttling is ineffective if you base it on IP address...

I didn't dictate any specific throttling algorithm. You are stabbing a strawman.

> an attacker obtaining the encrypted vault is probably not going to be able to decrypt many passwords,

That may not be how they breach them. It's an extra layer or device that may have an inadvertent security flaw. The more turtles in the stack, there more turtles there are to hack.

It will be interesting to see who gets rich after the bubble pops. Biggest Game of Financial Chicken Ever, Believe Me!

"Covfefe"

The random ones are too hard to remember, most will copy and paste. Either that, the help-desk is swamped with resets.

I'm not understanding why the traditional approach doesn't need throttling. Keep in mind a DOS attack is usually considered a smaller "sin" than a breach(es). If you allow too many retries, then the second sin is more likely. I see no third option*, it's either a DOS freeze or lots of retries.

If hackers find a design weakness in your company's preferred/required password-keeper, they can potentially hack them all. A company can allow multiple keeper brands, but then they either have to vet them all, or accept that some users will select a dodgy brand.

> I read your setup as a global throttle. If that's not what you meant...

* The best throttling and/or DOS defense strategy/algorithm is a more involve topic, but so far not a difference maker in what we are comparing.

> so now attackers can easily DoS your login system.

What keeps them for doing that with a traditional system? Even a traditional login screen should be throttled.

> Which is why you store it in a password-keeper

Another vector for hacking.

Correction: "With enough". Damn, I hope such simple typos are not a sign of Heimalzers, or whaddever itz called.

Without enough practice, many Alzheimer's patients can learn to get good enough.

The "requirements" for a secure passwords will keep trending up such that harassing users to write War and Peace to log in is a dead end.

The password server should be in a special box that throttles requests. It would have a very limited and primitive interface to the outside world; technicians would have to physically unlock it to service it. There would be a mirror server for a backup.

That way no hacker can run gajillion retries on a password without swiping the actual box.

This seems like a contradiction, I'm not following. Perhaps you mean "use" differently than I'm interpreting.

Schizophrenics are people too...including the orange guy who "sees" dog-eating Haitians tampering with voting machines using windmill radiation.

People with schizophrenia are usually considered "conscience" as long as they can usually dress themselves and eat on their own, so the bar is fairly low.

...than a big exercise in Laynes Law.

There is no consensus air-tight definition of "conscience", so enjoy your Never Ending Arguments.

> some come from other people.

Now I'll get sued for copyright infringement just for sleeping.