Does anyone make good SSDs any more?

[paulca]

I learnt the hard way that not all SSDs are the same.

I bought crucial SSDs for a virtualisation server.  They ended up in a gaming PC, where they work fine.

"I am told":  Most, average, consumer SSDs use a limited amount of high speed flash and back it with a much larger slow (to write) flash.  The result is that for short, bursty writes like a normal OS or gaming PC, it's absolutely fine.  However, if you hit it with transfers larger than it's high speed cache, it's performance will collapse back to that of the slow write flash.  This can be an order of magnitude slower or more.

This is before you consider the "write leveling" and other dynamic sectoring which needs maintenance cycles for various filesystems or performance tanks in much the same way as an old HD that was fragmented would.

All of this results in bottle necks causing bottle necks causing bottle necks.  When the transaction queue in the kernel is overflowing with filesystem write requests and some of them start getting told "No".  Things get unstable.  I mean who actually checked the return code for half those functions? (J/k)

Examples of how to trash a cheap Crucial SSD?  Do a "docker pull" on a project with 6 images built from 12 dependencies and docker launches 18 threads all writing Gb image files to disk.  This along caused a server wide lag and even a few kernel lock errors.

I was also told to... check for full capacity write performance tests.  If you spend twice as much you will find a few that can sustain full write speed for their full capacity.  Usually, lately the Samsung Evo range.

If you want some confidence in what you are getting, get your wallet out and go enterprise.

[mariush]

Correct, most SSDs switch a part of the free space to pseudo-SLC mode, storing only 1 bit in each cell instead of 3 bits for TLC or 4 bits for QLC - so for example, if you have 100 GB of free disk space, you may get around 32 GB of SLC write cache on TLC drives, or around 25 GB of SLC cache in the case of QLC drives.

The speeds drop from 3-7 GB/s down to around 500 MB to 1.5 GB/s on TLCs, while QLC drives will drop to around 100-250 MB/s when the write cache is full.

Drives like WD SN570 using older controllers will use smaller write caches, up to 13 GB, and speeds drop to around 600 MB/s when this amount is full....  https://www.techpowerup.com/ssd-specs/western-digital-sn570-1-tb.d467

The newer SN770  can pretty much switch the whole drive to SLC mode, and if you manage to fill it up, write speeds will drop to around 560 MB/s  ...  https://www.techpowerup.com/ssd-specs/western-digital-black-sn770-2-tb.d647

Samsung 980 I own can do up to around 122 GB of cache (the 1 TB version goes up to around 160) ... From 3.1 GB/s reads , 2.6 GB/s writes will go down to around 400 MB/s when the write cache is full.

The Samsung 980 Pro and 990 Pro can go up to around 250 GB of write cache but manage to flush to TLC memory much faster, at up to 1.5 GB/s

[coppice]

I have been burned by Western Digital too many times to trust them.  I was never able to confirm whether their SSDs support any form of power loss protection.

Western Digital bought Sandisk, and I assume that is the heart of their SSD business now. I haven't used much Western Digital branded SSD hardware, but I had terrible experiences with Sandisk reliability, from their SD cards to their SSDs.

[tom66]

One thing for sure: do not buy Kingston devices.

Kingston do not make their own flash, and the quality of their drives is consequentially variable from batch and model to model.  I have had both Kingston RAM and Kingston SSDs go bad well within normal hardware lifespan, and the fiasco over the V300 SSDs where Kingston shipped review samples with faster flash and controllers and then changed the production units has put me off for good.

[David Hess]

Kingston do not make their own flash, and the quality of their drives is consequentially variable from batch and model to model.  I have had both Kingston RAM and Kingston SSDs go bad well within normal hardware lifespan, and the fiasco over the V300 SSDs where Kingston shipped review samples with faster flash and controllers and then changed the production units has put me off for good.

When I started considering SSDs a couple years ago, I evaluated many manufacturers but quickly narrowed my list to those who make their own Flash, so Crucial and Samsung.  Intel was a close third.

[terrance]

YMTC produces some nice stuff. But seemingly they don't export much.

A Chinese youtube guy did a lot of accelerated aging and other torture tests (on samples from China's market). Some are still on going.

in case you are interested

https://www.youtube.com/@XFJDIY

Basically:

samsung is mostly ok-ish, but power consumption/temperature is high, thus PMIC tent to blow. if controller is dead, data recovery almost impossible
intel/hynix: high leakage, not suitable for off-line storage.
toshiba/kioxia/sandisk/WD: ok. data recovery almost impossible
ymtc: above ok, data recovery almost impossible for now, due to LDPC, maybe possible later.
cheap, OEM and BS "gaming" brands: varies ok to bad, but for some controllers, data recovery is likely possible.

[mariush]

Kingston do not make their own flash, and the quality of their drives is consequentially variable from batch and model to model.  I have had both Kingston RAM and Kingston SSDs go bad well within normal hardware lifespan, and the fiasco over the V300 SSDs where Kingston shipped review samples with faster flash and controllers and then changed the production units has put me off for good.

When I started considering SSDs a couple years ago, I evaluated many manufacturers but quickly narrowed my list to those who make their own Flash, so Crucial and Samsung.  Intel was a close third.

Crucial is Micron's brand. So it will only ever use Micron flash memory : https://www.techpowerup.com/ssd-specs/filter/?mfgr=Crucial

Kingston buys wafers from Micron and at least for ram, they bought them untested and did the validation in-house to save money.  I doubt they're doing it for nand.

See https://www.techpowerup.com/ssd-specs/filter/?mfgr=Kingston , pretty much Micron and Toshiba for everything.


Western Digital has a partnership with Toshiba and invested heavily into the Toshiba / Kyoxia nand factories. So technically it can be considered manufacturer.

example

2019 https://www.westerndigital.com/en-ap/company/newsroom/press-releases/2019/2019-05-16-toshiba-memory-and-western-digital-to-jointly-invest-in-flash-manufacturing-facility-in-kitakami-japan

Toshiba Memory Corporation and Western Digital Corp. (NASDAQ: WDC) have finalized a formal agreement to jointly invest in the “K1” manufacturing facility that Toshiba Memory is currently constructing in Kitakami, Iwate Prefecture, Japan.

2022 https://www.westerndigital.com/en-ap/company/newsroom/press-releases/2022/2022-04-14-kioxia-and-western-digital-jointly-invest-in-new-flash-memory-manufacturing-facility

Kioxia Corporation and Western Digital Corp. (NASDAQ: WDC) have finalized a formal agreement to jointly invest in the first phase of the Fab7 (Y7) manufacturing facility at Kioxia's industry-leading Yokkaichi Plant in the Mie Prefecture of Japan. With construction of the first phase of Y7 completed, the joint-venture investment will enable initial production output beginning in the fall of this year. This marks another important milestone in the 20-year strategic joint-venture partnership between the two companies.

This joint-venture investment adds a sixth flash memory manufacturing facility to the Yokkaichi Plant, enhancing its position as the world's largest flash memory manufacturing site. The first phase of the Y7 facility will produce 3D flash memory including 112- and 162-layer and future nodes.


2024, quite recent news   https://www.trendforce.com/news/2024/02/26/news-nand-flash-manufacturers-kioxia-and-wd-reportedly-set-to-resume-merger-talks-in-late-april/

NAND Flash Manufacturers Kioxia and WD Reportedly Set to Resume Merger Talks in Late April


and some interesting bits from that press release ... at Q3 2023, market share:

31.4% Samsung  31.4%
20.2% SK Hynix + Solidigm
16.9% WDC
14.5% Kioxia
12.5% Micron
04.6% Others


Anyways... if you're crying for Intel flash memory,  well the flash memory division was sold to SK Hynix ... they're using the Solidigm brand to sell drives with QLC memory that's based on the old Intel flash memories.

Solidigm had some SSDs that were very optimized for low power consumption, great for use in tablets and laptops.

[voltsandjolts]

There is no "the one".

Yes, I know. That was "my point".
Just curious as to why the OP is still banging on about this 8 months later, instead of just buying a half decent drive and getting on with other actually interesting stuff.

[BennoG]

For me the most important factor for an SSD is the quality of the wear-leveling.
And I will stay far away from the QLC type of drives. The problem is most manufactures don't mention type type of cell anymore.
As you can image if you have 16 different voltage levels in a single cell then even if a couple of electrons leak you have a different value (this can possible be corrected by ECC etc)

P.S. this is my personal opinion.

source of the image   https://www.kingston.com/en/blog/pc-performance/difference-between-slc-mlc-tlc-3d-nand

Benno

[paulca]

I think it stands that if you expect "one size fits all", you will find yourself out of pocket pretty quickly.

If you want an SSD for a desktop PC which will not be processing multi-gigabyte disk IO frequently, then that small (ish) burst of high speed is all you will need.  So a bog standard 50 quid a terrabyte is all you want.

If you, like I, want a SSD to survive under a busy virtualised environment running several servers and multi-gig IO transfers are common (docker/kurbenetes et al as well as transfering entire VMs around)... a 50quid a terrabyte drive will absolutely NOT do.

The largest bulks storage I have, accounts for about 95% of storage by volume, is "infrequent write", "frequent read". 

In cases where speed is important on read and less important than reliability, I strip 2 or more cheap SSDs together.  (consider a Steam drive or a drive you load games from).

[madires]

WD and SanDisk, today the same company for SSD production. Currently being sued for having produced external SSDs under both brand names which suddenly bricked and wiped the data on them. Also a scandal with critical firmware updates being needed for the SA510 line of WD branded internal drives.

And this one:
SanDisk Extreme Pro Failures Result From Design and Manufacturing Flaws, Says Data Recovery Firm: https://www.tomshardware.com/news/sandisk-extreme-pro-failures-are-due-to-design-flaw

So far all manufacturers had/have some issues. It's quite similar to things we have seen for classic spinning disks. Some product families are great, some need a firmware update, and some are junk. If your data is important to you make backups!

BTW, when buying an SSD make sure to buy from a reputable seller to reduce the risk of getting a counterfeit.

[WatchfulEye]

Having gone through a couple of dozen SSDs in the last few years, I have had a couple of early incidents of data loss (blocks of sectors are unreadable) - rather than total drive failure.

One was found, annoyingly, because a backup was failing - it turned out that the backup software would panic when it received a "bad sector" notification, and abort the entire backup with an unhelpful error message. Overwriting the affected sectors triggered a sector reallocation event (deleting the affected files, and then filling the drive's free space with random bits), and the drive was restored to service, and has worked fine ever since. I suspect that the data loss occurred within a couple of weeks of installing the drive, but the affected sectors were infrequently accessed, and a change in backup rotation resulted in a re-read (a scheduled full backup, rather than an incremental backup).

The other was found proactively during a periodic test-reading of the drive, and the affected files were restored from backup.

I've been trying a "burn in" procedure on new SSDs, perfoming a set of 5 full drive write/verify cycles, before putting the drives into service. So far, however, it hasn't yielded any results other than wasting time.

[coppice]

As you can image if you have 16 different voltage levels in a single cell then even if a couple of electrons leak you have a different value (this can possible be corrected by ECC etc)

You do realise they don't even try to recover 16 distinct levels from the flash cells, right? ECC on these things is not handled at the digital level. They use coding, and soft bit data recovery from the medium. This is similar to how most modern comms channels, and hard disks work.

[amyk]

For me the most important factor for an SSD is the quality of the wear-leveling.
And I will stay far away from the QLC type of drives. The problem is most manufactures don't mention type type of cell anymore.
As you can image if you have 16 different voltage levels in a single cell then even if a couple of electrons leak you have a different value (this can possible be corrected by ECC etc)

P.S. this is my personal opinion.

source of the image   https://www.kingston.com/en/blog/pc-performance/difference-between-slc-mlc-tlc-3d-nand

Benno

Those numbers should really be considered false advertising now. I don't think 100K SLC is even made anymore, and what parts are available are tiny. For example Micron MT29F32G08ABAAA is only 60K cycles. Samsung K9L8G08U0M (MLC) is 1.5K cycles.  Look at the datasheets that have escaped into the public, or divide "TBW" by the capacity of an SSD to get the real endurance figures. They've gotten a lot more secretive with TLC and QLC. They've also been slowly decreasing the retention to make the endurance numbers look better, since you can program for more cycles and make the cells leakier if you don't need the charge to stay as long. Those numbers above are 60K/10 years and 1.5K/5 years. TLC is more realistically a few hundred cycles for 5 years, and QLC probably around 100 for 3 years.

[David Hess]

I don't think 100K SLC is even made anymore, and what parts are available are tiny.

Swissbit sells 128G SLC drives ... for more than $1000 each.

Look at the datasheets that have escaped into the public, or divide "TBW" by the capacity of an SSD to get the real endurance figures.

That is the first thing I calculate.

TLC is more realistically a few hundred cycles for 5 years, and QLC probably around 100 for 3 years.

400 to 600 for TLC, and 200 for QLC but I only have one entry for those.  There was an Intel TLC drive that I considered with a calculated endurance of 3750, which is more like how MLC performs, so I do not know what was going on there.

[Infraviolet]

I opted for the WD Red SA500 in the end, no firmware updates released since production egan in 2019 does imply they've made no mistales they needed to fix. I'm doing my reinstall on to it today, so will soon see what its like.

[PlainName]

Bit late to this thread but my personal experience of WD NVMe SSD might be useful.

I have several disks in my main PC and currently a 1TB NVMe SSD as the boot disk with the following volumes on:

C: - Windows system and boot
D: - General data
E: - Applications (much less writing that the other volumes)
F: - Temporary stuff. Typically not much used
G: - Chip SDKs, IDEs and associated tools

After a bit of use (of the previous drive to the WDC) I was seeing some performance issues, but whenever I ran a speed test it came out with good numbers, until it didn't and then I invested in the WDC replacement. Around then I invested in HDTune Pro and did a benchmark of the drive in November 2021 (screeny attached). You can see that the early part of the drive is a bit slow,with a huge dip around 180GB in.

[Note: HDTune hits the disks sectors directly so is unaffected by filesystem and similar high-level abstractions.]

The volume layout has space between volumes for expansion, so C: might be 90GB allocated then 25GB free for expansion, D: has 75GB then 10GB free, etc. The upper 420GB of the disk is unallocated (for expansion and/or wear levelling). The early dip corresponds to drive C:, which is quite heavily accessed in both read and write modes.

OK so far. Now wind forward to August 2023 and see the second HDTune benchmark. The space used by drive data is clearly visible from the massive performance drop, and the unallocated space is still as good as new.

This isn't unique to WDC but seems to affect any NVMe SSD I've tried, some quite a bit worse than the WDC. I guess SATA SSDs have a low enough bus speed that the performance degradation doesn't affect transfer speeds.

[thm_w]

Can anyone give me a good reason not to select a WD RED SA500? (speed reasons are only "good" in my books if the speed of this would be actively worse than the spinning rust Toshiba MQ01ABD100 models I am used to)

I have been burned by Western Digital too many times to trust them.  I was never able to confirm whether their SSDs support any form of power loss protection.

I doubt any consumer grade SSDs would have usable power loss protection.
If you care you'd buy an enterprise grade drive (which has expensive supercaps inside), or, use a UPS (probably the cheaper option).

[coppice]

Can anyone give me a good reason not to select a WD RED SA500? (speed reasons are only "good" in my books if the speed of this would be actively worse than the spinning rust Toshiba MQ01ABD100 models I am used to)

I have been burned by Western Digital too many times to trust them.  I was never able to confirm whether their SSDs support any form of power loss protection.

I doubt any consumer grade SSDs would have usable power loss protection.
If you care you'd buy an enterprise grade drive (which has expensive supercaps inside), or, use a UPS (probably the cheaper option).

Some of the early SATA SSDs claimed to complete work in progress as the power fell. I don't know about current SATA drives, but NVMe drives don't seem to have anything big enough to store some energy to do that.

[David Hess]

Enterprise drives not only complete writes, but protect data which is in transit, so nothing should ever be lost.

At a minimum drives need to be able to complete write operations to prevent corrupting data, including the flash translation data structures.  In the past not all drives achieved this, but I hope all modern drives manage it.  Consumer drives advertised with power loss protection should be achieving this.

[WatchfulEye]

Some of the early SATA SSDs claimed to complete work in progress as the power fell. I don't know about current SATA drives, but NVMe drives don't seem to have anything big enough to store some energy to do that.

In the absence of capacitors for energy storage, drives can achieve power loss protection by ensuring that the specifications for write command semantics are strictly adhered to (i.e. only signalling that a write command is complete when the data is flash and metadata has been appropriately updated), and maintaining internal metadata integrity with journalling.  This does not necessarily mean that all in-progress writes will be committed, but it does mean that applications and file systems can maintain internal consistency, and databases can maintain transactional integrity/atomicity.

Enterprise NVMe drives tend to feature a few mF of tantalum polymer capacitors for power loss protection, allowing them to finish all work in progress, when power is removed. This means the drives can legitimately signal that data has been committed immediately on receipt of the write command - which can be important for overall system performance in transactional workloads.

[David Hess]

In the absence of capacitors for energy storage, drives can achieve power loss protection by ensuring that the specifications for write command semantics are strictly adhered to (i.e. only signalling that a write command is complete when the data is flash and metadata has been appropriately updated), and maintaining internal metadata integrity with journalling.  This does not necessarily mean that all in-progress writes will be committed, but it does mean that applications and file systems can maintain internal consistency, and databases can maintain transactional integrity/atomicity.

There are two things which make that difficult:

The Flash write procedure is controlled by a state machine which may glitch as the supply voltage drops, causing the write to complete at the wrong address, destroying other data.  Early SSDs suffered from this, but maybe the Flash chips have been improved to avoid this problem.

I do not know why they did this, but Flash devices which store multiple bits per cell can perform multiple writes to the same page.  If the second write does not complete, then the data from the first write is destroyed.  This is solved by writing all of the data for a page at the same time, but doubles (or triples, etc.) the minimum write size.

[amyk]

TLC SSDs are around $0.10/GB now. Where are the $0.30/GB SLC SSDs, which would use exactly the same amount of die area but won't need the additional cost of a more complex FTL/ECC and controller, while also lasting many times longer?

[PlainName]

Presumably no-one wants to pay three times more for a quarter of the capacity, so there is no profit in them.

[NiHaoMike]

The manufacturers should provide a way to reformat TLC or QLC as MLC or SLC, for those who want to trade off capacity for more endurance.