How Slow do TLC SSD Writes Really Get?

[Zarathustra[H]]

Hey all,

I am trying to diagnose some Samsung 850 EVO 1TB drives. I tried to find this in various reviews like Anadtech's, but I don't think they punished the drives as hard as I have.

What is the absolute worst case sequential performance you'd expect on one of these drives AFTER exhausting the SLC write cache in an environment where the drive will never see any TRIM commands?

Much obliged!

 

[Thevoid230]

So the closest thing in that review to what you describe would be "all LBA filled random write" testing. If you direct your attention to
https://www.anandtech.com/show/8747/samsung-ssd-850-evo-review/5
The first graph shows the result in write IOPS. Beyond that, you could look for an enterprise IOmeter write test.

For a description of the tests: https://www.anandtech.com/show/6433/intel-ssd-dc-s3700-200gb-review/3

 

[Maxx]

Depends on the drive. Generally speaking the worst-case is folding speeds which is about half of the native NAND speed, for example 80 MB/s on the 660p (which always folds). It also depends on capacity, for example with NVMe drives there will be "tiers" of write performance depending on the controller's saturation (channels and CEs or dies per channel). Typically this is 600 MB/s, 1200 MB/s, 1500 MB/s, with folding at 300 MB/s, 600 MB/s respectively to the first two. But this is not universal as it depends on the hardware generation (MT/s of the NAND) and many other things as well. The 850 EVO actually had multiple revisions (two or three controllers, three different NAND layers) but in general has good sequential performance regardless. The SLC cache ("TurboWrite") was small on the 850 EVO (12GB on the 1TB) and static so you're just hitting native TLC speeds which should be quite consistent. The controller is 8-channel with up to 8CE per channel but should be saturated even with 128Gb at just 8x4 which would be the 500GB (512GiB NAND) SKU. (later revisions had denser NAND but it was also faster)