older enterprise NVMe vs new consumer NVMe

[SnJ9MX]

With the latest FireCuda review, I am left wondering if my hunt for older enterprise U.2 NVMe on eBay, such as a 1.6TB Intel P4610 (datasheet here) is a bit of a fool's errand. I bought a GM7000 after reading a review here on STH and sure it is fast but I'm wondering how fast fast really is. I also bought a Cisco-branded WD SN200 (datasheet here) in April along with a carrier PCIe card.

To compare:

P4610 - 3200 MB/s R/W, 654k read IOPS, 220k write IOPS, 3 DWPD
SN200 - 3350 MB/s read, 2100 MB/s write, 835k read IOPS, 200k write IOPS, 3 DWPD
GM7000 - 7400 MB/s write (but I have it in a PCIe 3.0 slot, which limits it), 6400 MB/s read, 1000k r/w IOPS, 600 TBW (0.3 ish DWPD)

On paper the GM7000 is definitely faster, but for actual usage (scratch drive, database VMs, local LLMs, etc) the SN200 feels much faster. They're both in a machine with an Intel i7-8700k and 64 GB memory (running at 2666 MHz).

What's the verdict here? Is it worth pursuing older ebay drives (which are coming down nicely in price) or getting some of the newest stuff? Assume drive endurance isn't really a concern because in 99% of cases, it isn't.

 

[iGene]

At lease for database I will definitely recommend enterprise drives. fsync is heavy used in database while the performance for consumer drive after enabling fsync is miserable.

You can check this blog post from percona for more details

Why Consumer SSD Reviews are Useless for Database Performance Use Case

Reviews of consumer SSDs are not reliable when it comes to assessing a disk's suitability for database use. Percona CEO Peter Zaitsev explains how the use of fsync for database transaction commits can impact IO performance of SSD drives.

www.percona.com

 

[mrpasc]

The biggest difference between enterprise NVMEs and those „high level“ prosumer / gaming NVMEs is the fact, that the first one work even under sustained workloads according to their specs. Their specs are kind of „worst case“ specs. In opposite those shiny prosumer NVMEs are tested with ideal conditions (fresh, empty drive) with small workloads. Put some real heavy workload pressure on them and you will see: only marketing picture specs. They will not do well in real world enterprise workloads.
But: for a single user Homelab server they might be good enough, especially for single user workloads.

 

[nexox]

For anything with a lot of random writes, you want an enterprise drive with power loss protection, not to avoid data corruption when your cat pulls the plug out of the wall, but for performance. With PLP those fsyncs don't have to all fully write to NAND, the drive can return a successful fsync while the data (or sometimes more importantly, the internal metadata used to locate that data) is still in volatile storage. That means reduced write amplification and latency, on top of NAND that's usually more suited for that kind of workload to begin with, both of which are good for write performance, but also help prevent interference from writes causing outlier extreme read latency.

 

[i386]

One point for enterprise ssds: if you need guranteed power loss protection (plp) only enterprise/datacenter ssds offer that feature