Hacker News new | ask | show | jobs
by cwizou 1757 days ago
Yep this is terrible editorialising. It's faster on every scenario but very long sequential writes where, after 115 GB your performance drops from 2.5GB/s to below 1 GB/s. The older drives dropped earlier (40 GB) from 1.75 to 1.5 GB/s.

Toms article explains it but for those who want details, the SSD controllers reserves some amount of NAND cells for fast writes, what they call SLC cache. What that means is, instead of writing 3 bits per cell (in TLC), they only write one bit per cell on those NAND cells. They do that because writing one bit is much faster than writing 3.

So the controller maintains some "free" cells for fast write. In the BLU version, the 980 Pro controller that's used here has a different strategy which keeps more cells around (110ish GB vs 40 GB) for fast writes. As you fill up your disk, the controller will try to maintain this. You will see a drop when you go above that because the controller will start consolidating the data (the controller will also move data in it's idle time, consolidating the SLC writes to TLC writes).

This is why you do those specific tests as a reviewer, to analyse how the controller works, but it's not a real life scenario. A much more interesting test which we used to do back in the day was measuring performance when your disk is filled up (say 80, 90%). This is usually where you see those SLC cache strategy no longer working (depending on how good the controller is, some just don't consolidate data and you end up having poor performances).

Anyway, it's a pretty irrelevant change for 99% of users and the SSD will be faster on all other scenarios.

Considering they did a hardware revision, updated spec sheets and new packaging, and changed to a better controller (from the 980 Pro), this is nothing like what WD did.

1 comments

As someone who actually uses the storage that I purchase, I strongly disagree with the suggestion that this is a largely-irrelevant change. Few if any users only write to empty disks.

If my experience with other QLC drives is instructive — and I hope it's not — I suspect the "runway" you referenced is minimal if you start writing with the disk at, say, 25% capacity.

I took a few shortcuts in my explanation and may have implied that the disk full case was similar to what was shown here (it's not completely related). All in all this is really controller dependant. Some controllers do the bare minimum to show good speeds in benchmarks, which means they never do much consolidating once every cell has been written once. This used to be the norm when SSDs were relatively new. 3 years ago this wasn't the case anymore, but for the most bargain brands (of controllers, you may be disappointed by "good" brands using poor controllers, Intel for example).

At 25% capacity you usually don't see any degrading in performance on the best controllers (I would put Samsung in there). QLC makes things harder on controllers (4 bits) but unless you are in a terrible spot (say you just filled 80%, and try to fill more without leaving enough time to consolidate), you are rarely doing terrible.

Which QLC drive did you have a bad experience with ? Samsung's QLC drive as far as I know recover their "SLC" cache pretty much always for example, according to Toms in 5 minutes of idle time [1]. And this is with their cheapest (I think) model out there.

[1] : https://www.tomshardware.com/reviews/samsung-870-qvo-sata-ss...