SSD Thread: (its good, ignore the anime)

The Something Awful Forums > Discussion > Serious Hardware/Software Crap > SSD Thread: (its good, ignore the anime)

«‹›221 »

Klyith: Aug 3, 2007; GBS Pledge Week

mikey posted:

What little I can find from user reports with the new hardware suggests that the write performance tanks badly as the drive passes 50-75% full, which is definitely not the case for the MP44 or NM790. Is that a factor of using QLC flash that I've never heard of before?

QLC drives typically use almost all their free space for pSLC cache to hide the terribly slow QLC writes. If a 1TB drive is 75% full, the pSLC cache will be small enough that you'll start having noticeable QLC penalty doing "normal" things like installing a 60gb game (or game patch). So I would guess that these reports are people who bought the smaller drives and didn't know they were getting QLC.

On a 4TB drive, being 75% full means you still have ~250GB of pSLC, so your cache will cover your regular tasks.

However, buying a drive that you can only fill 75% full instead of 95% full kinda sucks, and means you're not really getting a deal on the price/GB. That's why I generally think QLC is meh for anything but true bulk storage (or desktop-apps-and-browser users). However, if your games library doesn't have a lot of turnover you might not notice.

mikey posted:

Do I try the MP44 for $20 more even though it seems to have a relatively high failure rate, or go with the Chinese company wearing the skin of a former US brand Buffalo Bill-style, because their drive has seemingly better performance and reliability than similar options and adequate support at this point?

Personally I'd return it on principle for being bait-switched.

The MP44 and the Lexar NM790 are basically the same thing. Same controller, they're both going to be using whatever YTMC flash was cheap the week they were made. If you're seeing a performance difference, make sure you aren't looking at reviews with different capacity. It looks to me like Lexar sent everyone 4TB drives to review, while the MP44 reviews are 2TB.

Teamgroup actually has a bigger drive write warranty. I honestly doubt there's really any reliability difference -- I feel like if you can see more failure reports for the MP44, it might just be more people bought them.

# ? Apr 8, 2025 14:25

Adbot: ADBOT LOVES YOU

# ? May 23, 2025 00:25

Shumagorath: Jun 6, 2001

Klyith posted:

QLC drives typically use almost all their free space for pSLC cache to hide the terribly slow QLC writes.

My ADATA write speeds would oscillate well south of that, and Samsung QVO falls off a cliff at 71GB per 4TB from out-of-box, so I don't know if that was always the case?

# ? Apr 8, 2025 22:14

mikey: Sep 22, 2002

Klyith posted:

QLC drives typically use almost all their free space for pSLC cache to hide the terribly slow QLC writes. If a 1TB drive is 75% full, the pSLC cache will be small enough that you'll start having noticeable QLC penalty doing "normal" things like installing a 60gb game (or game patch). So I would guess that these reports are people who bought the smaller drives and didn't know they were getting QLC.

On a 4TB drive, being 75% full means you still have ~250GB of pSLC, so your cache will cover your regular tasks.

However, buying a drive that you can only fill 75% full instead of 95% full kinda sucks, and means you're not really getting a deal on the price/GB. That's why I generally think QLC is meh for anything but true bulk storage (or desktop-apps-and-browser users). However, if your games library doesn't have a lot of turnover you might not notice.

Aha, that makes sense. I understood how pSLC worked, but didn't make the connection that it's both limited by free drive space, and that QLC:pSLC is 1:4 instead of the usual 1:3, so you get even less functional cache vs TLC as the drive fills up, in addition to worse post-cache speeds. And while the direct-to-QLC speeds have improved, at 4TB they're still like 150MB/s vs 2500MB/s for TLC, which is just not worth saving $20-30.

Klyith posted:

Personally I'd return it on principle for being bait-switched.

The MP44 and the Lexar NM790 are basically the same thing. Same controller, they're both going to be using whatever YTMC flash was cheap the week they were made. If you're seeing a performance difference, make sure you aren't looking at reviews with different capacity. It looks to me like Lexar sent everyone 4TB drives to review, while the MP44 reviews are 2TB.

Teamgroup actually has a bigger drive write warranty. I honestly doubt there's really any reliability difference -- I feel like if you can see more failure reports for the MP44, it might just be more people bought them.

Yeah, I returned it. And the price has gone up $10 in the last day, lol. First and last Silicon Power purchase for me, and now I feel bad because I recommended the 1-2TB US75 to one or two people, never even considering that those could have been QLC-swapped. e: TPU still lists the 500GB-2TB models as TLC, but people are already reporting QLC down to the 500GB level.

TPU reviewed the Lexar at 2TB and 4TB and they were both pretty performant in real-world stuff, even if they had some weak spots. But I'll probably go with the Evo Plus just out of paranoia now. :negative:

I never pay too much heed to the warranty durability, maybe there's a difference by binning or something, but probably not as much as the mfg is claiming. It seems like you can generally assume 600-800x for TLC, and 300-400x for newer QLC, and most drives will happily keep writing far past that, but that's roughly where the failures start to appear for some samples whenever some site does write-destruct testing on a batch of drives.

Shumagorath posted:

My ADATA write speeds would oscillate well south of that, and Samsung QVO falls off a cliff at 71GB per 4TB from out-of-box, so I don't know if that was always the case?

It's more a thing for newer QLC drives with better controllers trying to masquerade as 'real' drives that you might actually be able to put an OS on. And they mostly succeed - the MP44Q and NQ790 are much faster overall than most pre-2022 drives right up until you exceed their available cache.

mikey fucked around with this message at 23:03 on Apr 8, 2025

# ? Apr 8, 2025 22:57

Klyith: Aug 3, 2007; GBS Pledge Week

Shumagorath posted:

My ADATA write speeds would oscillate well south of that, and Samsung QVO falls off a cliff at 71GB per 4TB from out-of-box, so I don't know if that was always the case?

Yeah it's not always been a thing. For the first couple QLC models they still used static size caches. And then the first drives with dynamic cache size were afraid to go whole hog. I think the Rocket Q was the first one that did "gently caress it, use the whole drive for cache"? But it's been an increasingly popular thing for a while now. If you aren't buying something old like a QVO it's very likely to be the case now. Even TLC drives have gotten in on the action too!

It makes the hit when you run out of cache from a large write even worse, because you have to collapse pSLC data to your MLC at the same time as writing new data. But on the flip side it's a generally good thing for day-to-day use, because pSLC writes don't damage endurance as much. Aggressively committing to pSLC and waiting longer to write to the MLC is good for most desktop use.

edit:

mikey posted:

I never pay too much heed to the warranty durability, maybe there's a difference by binning or something, but probably not as much as the mfg is claiming. It seems like you can generally assume 600-800x for TLC, and 300-400x for newer QLC, and most drives will happily keep writing far past that, but that's roughly where the failures start to appear for some samples whenever some site does write-destruct testing on a batch of drives.

Yeah I don't think it makes a lick of difference, as I say I think those two drives are gonna be near-identical. Was just a thing that teamgroup has a marginally more confident warranty than lexar.

Klyith fucked around with this message at 23:34 on Apr 8, 2025

# ? Apr 8, 2025 23:30

WhyteRyce: Dec 30, 2001

I've heard that 2025 is the year of QLC so get on board assholes

# ? Apr 9, 2025 18:12

orcane: Jun 13, 2012; Fun Shoe

WhyteRyce posted:

I've heard that 2025 is the year of QLC so get on board assholes

no thx

# ? Apr 9, 2025 23:41

HalloKitty: Sep 30, 2005; Adjust the bass and let the Alpine blast

orcane posted:

no thx

Yeah, I still see tlc as kind of janky

# ? Apr 10, 2025 12:04

mikey: Sep 22, 2002

PCPartPicker just overhauled their benchmark methodology and database. They added a bunch more drives, multiple queue depths for random and sustained read/write, with all tests other than drive-fill done at 50% full. Still haven't fixed the absent DRAM cache value for the majority of DRAM drives though.

Naturally, they included the 4 TB 990 Evo Plus exactly one day after I bought one. Nobody else had benchmarked it, and I was hoping it would have better random read than the 2 TB, but nope, of course it's worse. At least it has better sequentials and random write than the others I was looking at, because the random read is worse than some QLC drives. Why are you like this, Samsung.

# ? Apr 12, 2025 07:28