I'd stripe 2-3 of them. Does the TLC flash achieve anywhere near the PCIe 4 interface speed after the buffer is exhausted in a few seconds?
I've found the current drives are crapping out well below 2000MB/sec. sustained writes, basically not better than the old MLC 970 Pro. I think enterprise grade U2 or other interfaces are needed to sustain high writes.
According to Samsung's marketing in the link you posted it gets pretty close to the PCIe4 limit but to your point, this could be a burst speed versus sustained. We'll have to wait and see what the in-depth reviews come up.
PCIe 4.0’s highest theoretical sequential read speed is 8000 MB/s — 990 PRO reaches 7,450 MB/s as of Q3 2022.
The U.2 is just the physical interface, it doesn't offer any more theoretical bandwidth when compared to M.2 since it's just the connecting physical interface to PCIe bus where the same NVMe protocol comes into the picture. Those enterprise drives just tend to have larger amounts of over-provisioning NAND, better controllers, and maybe more onboard RAM with power loss protection.
Yes of course they are better due to the enterprise needs. I also suspect that the capability of using a lot more power due to the connectors and dissipation from metal enclosures makes a difference.
Sustained writes are always a problem for me. The trend lately is for many SSDs to be slower than the previous ones. It's sneaky and a type of manufacturing cost reduction. Hiding behind the pseudo-SLC buffer does no good when you need to move a TB or two to the SSD and it slows down. Only a few websites test SSDs in the dirty state where the sustainable writes are shown.
This might be a use case where you consider an enterprise level SSD to help meet your performance needs. You can get adapters to convert U.2 to PCIe or M.2 slots and use a 2.5" SSD. I do this in a couple of my systems right now.
Another alternative is to buy one of these newer 990 Pro when it comes out and test it yourself. If it sucks, return it.
When we qualify enterprise SSDs at work, we pre-fill them with random non-compressible data before beating on them with FIO utility. We had the same concerns with sustained performance when the drive has been filled and we perform a lot of writes. Since getting true random data from linux under /dev/urandom can be painfully slow, we leveraged this neat trick to use openssl to assist when generating lots of random data quickly.
I use Intel P4500s for continuous backup and database dumps. They are TLC drives that top out at around 3GB/sec writes but they're affordable and use a convenient 2.5" form factor. I'm sure I have at least a few dozen in service. They've all been extremely well behaved.