News AMD Engineers Show Off 'Infinitely' Stackable AM5 Chipset Cards

[InvalidError]

No, that's why we're talking about NAND/Optane, rather than DRAM. Unlike DRAM, they won't lose the intermediate copy during power-loss.

NAND may not disappear on shutdown but it is known to get corrupted on unexpected shutdown too. If literally every write goes through NAND, you will burn through the write cycles pretty quickly unless you make the NAND so large that you might as well just have an SSD.

 

[bit_user]

NAND may not disappear on shutdown but it is known to get corrupted on unexpected shutdown too.

We're talking about a properly-engineered solution. So, no, that wouldn't be an issue.

If literally every write goes through NAND, you will burn through the write cycles pretty quickly

Yes, you already made that point. I underestimated the endurance issue. @Kamen Rider Blade is right that Optane would probably be needed for it to be viable.

 

[InvalidError]

We're talking about a properly-engineered solution. So, no, that wouldn't be an issue.

A "properly engineered solution" would start with redundant PSUs each with its own UPS to ensure power failures cannot randomly kill the OS since the OS getting interrupted mid-write is the first step where things can go horribly wrong.

 

[bit_user]

A "properly engineered solution" would start with redundant PSUs each with its own UPS

You're being hyperbolic. There are SSDs with built-in capacitors which ensure they have enough power to dump the contents of their DRAM buffers to NAND. It's a standard feature of datacenter SSDs, but also something you can find in some upper-tier consumer models. I'm not sure if any M.2 drives have them, but some SATA models do/did.

to ensure power failures cannot randomly kill the OS since the OS getting interrupted mid-write is the first step where things can go horribly wrong.

This is why we have journaling filesystems and why they use write barriers to limit I/O reordering.

 

[InvalidError]

You're being hyperbolic. There are SSDs with built-in capacitors which ensure they have enough power to dump the contents of their DRAM buffers to NAND. It's a standard feature of datacenter SSDs, but also something you can find in some upper-tier consumer models. I'm not sure if any M.2 drives have them, but some SATA models do/did.


This is why we have journaling filesystems and why they use write barriers to limit I/O reordering.

All of that ON TOP of having UPSes as the first line of defence because no amount of storage system ruggedness does you any good against losing everything that is still floating around in system RAM.

 

[bit_user]

All of that ON TOP of having UPSes as the first line of defence because no amount of storage system ruggedness does you any good against losing everything that is still floating around in system RAM.

No, not really. Correct use of a Journalling filesystem (i.e. without disabling barriers) will even offer protection in the event of kernel panics, which is equivalent to an unmitigated power loss.

However, journalling just protects against filesystem corruption. If applications' I/O is designed correctly, it will also protect against application-level data corruption, but that's incumbent on the app developer. At this point, we're not talking about hardware any more.

A UPS lets you schedule an orderly shutdown, in which case you don't even need journalling filesystems or power-loss protection in your storage devices. ...except for the odd kernel panic, or the clumsy employee who unplugs the wrong machine.