Tom YumQLC drives also have pretty ordinary write speeds after the SLC cache is exhausted/drive fills up. With modern games requiring 10's of GB of data to be written during install, QLC has real world shortfalls.

You think your internet connection is fast enough for that to matter? Because it very likely isn't. And even if it is, where you are downloading it from likely isn't. Oh, and then there is the fact that the 1TB+ QLC drives all have 200GB+ SLC caches.

newtekie1Yes it is correct that a system/boot drive is a high write scenario, but even that will not wear out a QLC drive.

Nonsense. QLC P/E endurance is measured in the low to mid hundreds(600 at best). This is compared to TLC's 6500 or 7000 currently. That's less than 10% durability comparitively. Sorry, the numbers just do not support your statement. By the numbers, QLC is not suitable for boot drives or any scenario where frequent drive writes take place. This is physics, not opinion.

There's nothing wrong with QLC as a technology IMO. Yeah, QLC is slower and has less endurance than TLC. However, it is cheaper to manufacture, which is a great advantage.

However, if the main advantage to a technology is "cheaper to manufacture", then it better be "cheaper to purchase", or else I'm going to laugh them off stage. $120 is the going price for 1TB TLC right now, so 1TB QLC better be ~$100 before I even consider it.

Because QLC is an inferior product (that... to be fair... uses ~33% fewer resources to make), I expect ~33% lower price on the item. If not 33%, then at least some kind of a discount compared to TLC technologies. I know how the market works, I recognize that companies want to raise prices and they're in the right to do so. But as a consumer, I'm also in a position to ignore poor price-strategies, as well as recommend to others to do the same.

-------------

It was like those DRAM-free SSDs a few years ago.I know Toshiba made a few: but cutting out DRAM out of the drive entirely, they saved on parts and made a cheaper-to-manufacture SSD. Unfortunately, the darn SSD wasn't lower in price for some reason, so there's not much to do aside from laugh at the offer...

dragontamer5788However, it is cheaper to manufacture, which is a great advantage.

Only if the end user does not care about drive longevity. The vast majority do. Cost where it comes to storage must be balanced with durability. Hell, I would recommend a mechanical HDD long before I'd ever suggest a QLC based drive. All day, every day.

lexluthermiesterNonsense. QLC P/E endurance is measured in the low to mid hundreds(600 at best). This is compared to TLC's 6500 or 7000 currently. That's less than 10% durability comparitively. Sorry, the numbers just do not support your statement. By the numbers, QLC is not suitable for boot drives or any scenario where frequent drive writes take place. This is physics, not opinion.

The endurance numbers don't support your opinion. A typical system drive sees less than 6TBW per year. Even a shitty 40TWB drive is going to last more than 5 years. These are facts. Sorry, your opinions just don't align with reality.

newtekie1The endurance numbers don't support your opinion.

Then you seem to not understand the numbers.

newtekie1A typical system drive sees less than 6TBW per year. Even a shitty 40TWB drive is going to last more than 5 years.

That's an assumption on your part. Of the first 20 drives I installed based on QLC, 6 of them came back defective less than a year later. They were a brand name model. Further testing showed wear-leveling was very much less than acceptable. The company refused to honor the warranty stating that the drives must have been "abused", which was complete & utter twaddle. Even some of the latest drives do not pass testing in my shop. No QLC drive will be allowed to be installed/sold in my shop. Experience is the basis of my perspective. You can believe whatever you wish.

lexluthermiesterOnly if the end user does not care about drive longevity. The vast majority do. Cost where it comes to storage must be balanced with durability. Hell, I would recommend a mechanical HDD long before I'd ever suggest a QLC based drive. All day, every day.

You're right in that endurance is important, but only to a point.

SSDs may lose half their endurance whenever they go up these "halvenings" (SLC -> MLC -> TLC and now -> QLC). However, the easiest way to "restore" that lost endurance is to simply double the capacity once more.

A 1TB TLC will have 600 TBW endurance (with current technology), a 2TB QLC will also be nearly 600TBW (1/2 from the TLC->QLC change, but then 2x from simply having more space to load-balance the endurance across). As such, as SSD capacity increases, we reach a point where (eventually), the consumer is unlikely to ever write that amount over say... 5 years or 10 years (or whatever the expected lifetime of the drive is). As such, QLC will eventually become the defacto standard (assuming storage sizes keep increasing). I don't know if its at 1TB, 2TB or at 4TB, but somewhere in there, QLC endurance will be reasonable.

There are also endurance advancements every now and then: VNAND and similar characteristics. With the right process technology leading to superior "natural endurance", maybe there won't even be a need to increase the capacity / load balance for endurance purposes.

-----------

I wouldn't buy QLC today. But who knows what next year will bring? I can imagine a future QLC drive that I'd buy.

dragontamer5788You're right in that endurance is important, but only to a point.

I disagree. Durability is always important.

dragontamer5788I wouldn't buy QLC today. But who knows what next year will bring? I can imagine a future QLC drive that I'd buy.

It will take a big technological or electrochemical advancement to make NAND more durable as it pushes PLC/HLC/ETC limits. Anything beyond QLC currently has no expectations of marketable durability, controller based wear-leveling schemes or not. QLC is only barely so and that requires skilled marketing nonsense(read advertising deceptions) and a certain level of gullibility of the buying public.

lexluthermiesterI disagree. Durability is always important.

Clearly not! Because we've already moved off of SLC and MLC drives... drives which are 4x and 2x more durable than TLC drives. The current market has spoken: TLC is the sweet spot today. But I doubt it will remain the sweet-spot forever.

lexluthermiesterAnything beyond QLC currently has no expectations of marketable durability, controller based wear-leveling schemes or not. QLC is only barely so and that requires skilled marketing nonsense(read advertising deceptions) and a certain level of gullibility of the buying public.

I dunno. I've learned a lot about error correction back in my college years, and I've revisited information theory (Shannon limits) and whatever. I don't know what kind of error-correction they're doing on these drives, but there's a spectacular amount of ECC theory that can correct all sorts of errors (LDPC, Turbocodes, Reed Solomon, etc. etc.).

The most interesting, with regards to QLC / PLC / etc. etc. is "soft decoding" (which I believe applies to LDPC and Vertibi). Wherein you can take a sequence of "soft vectors" and calculate the average distance expected... anyway... I'm not really an expert on this subject. But... I've seen some crazy ass algorithms out there. Someone out there might be able to solve these error-correction problems that are now occurring on QLC.

The real question is: what is the Shannon limit of NAND-flash. Yeah, we're today thinking of it as "4-bits per cell" vs "3-bits per cell", but any information theorist who has studied gaussian-channels (aka: AWGN) knows that information theory goes far, far, far beyond just 4-bits per cell. In fact: AWGN is about extracting the maximum information out of gaussian analog channels ("infinite bits" so to speak: 1.025 volts vs 1.023 volts could change the information). With the correct algorithms, its possible to in fact reliably pull data up to the Shannon limit of any channel.

Case in point: WiFi6 is 10-bits per radio signal, aka QAM1024. Some crazy information-theory math exists out there! I'm frankly not convinced that we've hit the sweet spot of the error-correction technologies applied to NAND Flash.

------------

Anyway, I don't really know how this stuff works. But... that's the kind of math that goes through my head when I think about it. I really have seen some crazy algorithms that fix all sorts of weird situations and error-cases. Its almost surprising how strong modern day error-correcting codes have become. As such, progress must move forward. There's simply hasn't been enough time to figure out the limits of NAND Flash yet IMO.

lexluthermiesterThat's an assumption on your part. Of the first 20 drives I installed based on QLC, 6 of them came back defective less than a year later. They were a brand name model. Further testing showed wear-leveling was very much less than acceptable. The company refused to honor the warranty stating that the drives must have been "abused", which was complete & utter twaddle. Even some of the latest drives do not pass testing in my shop. No QLC drive will be allowed to be installed/sold in my shop. Experience is the basis of my perspective. You can believe whatever you wish.

If the TBW stat on the drive was less than the rated endurance on the warranty then whatever brand this drive was in the wrong. Period.

$154 for 1TB QLC is tough sell, there's plenty sub $100 drive served same purpose and durability.

dragontamer5788Clearly not! Because we've already moved off of SLC and MLC drives...

Those are choices by manufacturers, not by the public. If they offered us SLC and MLC drives in usage capacities, we would buy them. But makers want profits not what is good...

dragontamer5788there's a spectacular amount of ECC theory that can correct all sorts of errors (LDPC, Turbocodes, Reed Solomon, etc. etc.).

But none of that can correct physical electrochemical failure, which ALL types of NAND are susceptible to, QLC much more than TLC, MLC and SLC. The question is not whether a NAND cell will fail, but rather the question is when. Eventually all of it WILL fail, it is a mathematical certainty. Again, physics wins the debate.

dragontamer5788The real question is: what is the Shannon limit of NAND-flash.

You are talking about the "Claude Shannon" theory? Not sure how you are applying that to the science of NAND flash storage as they are unrelated principles. NAND flash is a chemical phase change storage principle whereas the "Shannon limit" applies to RF error correction, IIRC. Regardless, there are limits to how many times the phase of NAND cell can be changed. So unless a new, more durable chemistry is developed, QLC NAND will always be dubious at best as a frequent use storage medium, to say nothing of PLC or HLC...

newtekie1If the TBW stat on the drive was less than the rated endurance on the warranty then whatever brand this drive was in the wrong. Period.

I will not disagree with you, they were in the wrong.

1d10t$154 for 1TB QLC is tough sell

Whole heartedly agree with you. I would say and impossible sell. Might be good if it were TLC and the drive had a beefy DRAM cache...

lexluthermiesterYou are talking about the "Claude Shannon" theory? Not sure how you are applying that to the science of NAND flash storage as they are unrelated principles. NAND flash is a chemical phase change storage principle whereas the "Shannon limit" applies to RF error correction, IIRC.

There's a reason why CD-ROM and DVD-ROM drives have error-correction codes based off of Shannon coding theory: Because the math applies directly to storage, just as much as it does to RF signals.

en.wikipedia.org/wiki/Cross-interleaved_Reed%E2%80%93Solomon_coding

That's the algorithm for CD-ROMs. Reed-Solomon codes, which were popular in from the 1970s through the early 2000s (though better algorithms exist today). These are fully within the realm of Shannon coding theory. It seems that DVDs and BluRays also used these codes, based on my search on Wikipedia.

All storage works as a communication channel: its just between you and yourself, instead of between Alice and Bob. As such, AWGN, Binary Symmetric Channels, and all that theory from the RF-world applies just as well to storage as it does RF. After all: all storage is is a "delayed communication", if you really think about it.

newtekie1You think your internet connection is fast enough for that to matter? Because it very likely isn't. And even if it is, where you are downloading it from likely isn't. Oh, and then there is the fact that the 1TB+ QLC drives all have 200GB+ SLC caches.

What does internet speed have to do with exactly anything here? I download stuff straight to my NAS and then pull it from there over a 10Gbps connection. Sure, I'm still not likely to max out even a QLC drive, but your argument makes no sense. I also copy data between my SSDs at times, as yes, some of us have multiple SSDs in our systems.

lexluthermiesterNonsense. QLC P/E endurance is measured in the low to mid hundreds(600 at best). This is compared to TLC's 6500 or 7000 currently. That's less than 10% durability comparitively. Sorry, the numbers just do not support your statement. By the numbers, QLC is not suitable for boot drives or any scenario where frequent drive writes take place. This is physics, not opinion.

Intel claims 370 P/E cycles or 0.2 DWPD for these drives...

@dragontamer5788 you're making some very flawed assumptions with regards to NAND flash durability. It's not a halving between each technology transition, it's actually much worse, in most cases. SLC offers 50-100k P/E cycles. MLC brought this down to 10-20k depending on the type of MLC. Planar TLC was almost as bad as QLC at 500-1,000 P/E cycles, whereas 3D TLC brought it back up to around 3k. Intel claims 370 P/E cycles for this specific drive, which is apparently 85% more P/E cycles than the 660p.
So how is any of this halving? In some terms, it's a disaster, but luckily for most people, they read a lot more data from their drives then they write, so for consumer use 3D TLC is still acceptable when you look at the price/performance.

TheLostSwedeWhat does internet speed have to do with exactly anything here? I download stuff straight to my NAS and then pull it from there over a 10Gbps connection. Sure, I'm still not likely to max out even a QLC drive, but your argument makes no sense. I also copy data between my SSDs at times, as yes, some of us have multiple SSDs in our systems.

Because most of the legal ways to get those big games don't allow you to do that. They have you stream the data straight to whatever computer you are installing them on though whatever distribution client the game is available on.

And why download straight to a NAS if you are installing the game on a computer? You are still waiting on your internet connection to bring you the data so the internet connection is always the speed bottleneck. Maybe if you are installing the game on multiple computers, but how many people do that? And as you said, you won't hit the QLC anyway thanks to SLC caching, so your whole point about games was wrong from the beginning.

newtekie1Because most of the legal ways to get those big games don't allow you to do that. They have you stream the data straight to whatever computer you are installing them on though whatever distribution client the game is available on.

And why download straight to a NAS if you are installing the game on a computer? You are still waiting on your internet connection to bring you the data so the internet connection is always the speed bottleneck. Maybe if you are installing the game on multiple computers, but how many people do that? And as you said, you won't hit the QLC anyway thanks to SLC caching, so your whole point about games was wrong from the beginning.

Eh? Where did I mention games? Please, tell me...
Now can I have some of whatever you're smoking?

newtekie1Oh, and then there is the fact that the 1TB+ QLC drives all have 200GB+ SLC caches.

Where in "all have 200+ GB SLC caches" do the ones with dynamic SLC caches (that decrease in size as the drive is filled) go?

This doesn't look like 200+ GB to me:

Source: www.anandtech.com/show/13078/the-intel-ssd-660p-ssd-review-qlc-nand-arrives

dragontamer5788Yeah, that's a QLC drive. You can find a proper well-reviewed TLC drive 1TB like HP EX920 for $115: www.newegg.com/hp-ex920-1tb/p/N82E16820326778 (which happens to be a winner in Random QD1 tests).

You can find a brand-name SATA TLC drive for $90 btw. Such as Crucial MX500 (www.newegg.com/crucial-bx500-1tb/p/N82E16820156231).

If you want cheap/capacity, SATA-TLC for the win. If you want performance, NVMe TLC for the win. Why buy the low-endurance low-performance QLC product in the middle? Once you account for QD1 performance, SATA-TLC starts to make more and more sense for cheap capacity.



And the HP EX 920 has 3500 MBps read and 650TBW, because TLC beats QLC in both speed and endurance. Furthermore, actual practical tests (aka: QD1 IOPS speeds) strongly favor TLC drives. So the "hidden" advantage leans towards TLC drives like HP EX920 or even SATA-TLC drives in practice. Especially when the very slow continuous write speeds of QLC are taken into account ("continuous write" meaning after the SLC cache is exhausted).

dragontamer5788I disagree.

The HP EX850 series was available for years, before being upgraded to the HP EX920 recently. HP's TLC drives have always been rated with high-reviews and "practical tests". Furthermore, Crucial's offerings (MX500) was following in a long line of TLC drives as well: MX100, BX100, etc. etc.

Being interested in both of your recommendations for the HP 920 drive, I went looking on Amazon.

It appears I can get an HP 920 1tb for $125 on amazon. But, I also see there is an HP 950 drive which is quoted as being quite a bit faster for the same money.

BTW, the 920 appears to be quoted at 3200/1800 while the 950 is quoted at 3500/2900.

TheLostSwedeEh? Where did I mention games? Please, tell me...
Now can I have some of whatever you're smoking?

Follow the chain of posts. Here I'll help you:

Tom Yum said: SLC cache doesn't help because game installs are too big.
I Responded to Ton Yum: But you're downloading those games from the internet, which is still slower than QLC.
Then you Responded to me: What does internet speed have to do with anything?

You decided to chime into a conversation about games and then want to act like we aren't talking specifically about games. Sorry, we were talking specifically about installing games. If you can't follow the conversation, don't chime in.

milewski1015Where in "all have 200+ GB SLC caches" do the ones with dynamic SLC caches (that decrease in size as the drive is filled) go?

This doesn't look like 200+ GB to me:

Source: www.anandtech.com/show/13078/the-intel-ssd-660p-ssd-review-qlc-nand-arrives

Yeah, thats the oldest generation which isn't produced anymore. It was replaced by the 665p which has 200GB SLC cache for the 1TB model. As do most other QLC drives now.

newtekie1Yeah, thats the oldest generation which isn't produced anymore. It was replaced by the 665p which has 200GB SLC cache for the 1TB model. As do most other QLC drives now.

Actually it's the 660p that is still available. The 665p was phased out.

lexluthermiesterThat's an assumption on your part. Of the first 20 drives I installed based on QLC, 6 of them came back defective less than a year later. They were a brand name model. Further testing showed wear-leveling was very much less than acceptable. The company refused to honor the warranty stating that the drives must have been "abused", which was complete & utter twaddle. Even some of the latest drives do not pass testing in my shop. No QLC drive will be allowed to be installed/sold in my shop. Experience is the basis of my perspective. You can believe whatever you wish.

In the places I've worked, people seem to "burn through" SSDs like crazy. I've seen many cases where machines get unstable due to file corruption, caused by defective SSDs(SMART errors), often after about 2-3 years of usage (in a few cases 1 year), and all of these machines have only been used for programming/compilation and light web browsing. Yet the few SSDs that did show statistics was way past their endurance rating, dozens of TB written, and it makes no sense because I know what the users have been doing. My theory is that somehow writing millions of tiny files causes the SSD to write larger cells plus the wear leveling going crazy, so 1 GB/day of files turns into "200 GB" worth of wear. I've been building computers since the 90s, and this just feels off to me, we seem to be approaching the reliability of floppies.

If my experiences are representative for people doing a lot of small writes quickly wearing out SSDs, then these endurance ratings are probably as misleading as the MTBF ratings on HDDs. I actually find it hard to find solid solutions when the storage mediums are so unreliable. And of course I have backups, that's not the problem, the problem is the time wasted on recovery and cleaning up etc. A couple of years ago I moved to a separate OS SSD on my desktops and my file server, and making the OS drive "disposable" (meaning it contains nothing important), if something goes bad I can format it (or replace it), reinstall the OS and run a few scripts and I'm back up in 15-30 min. In my home workstation, my workspace resides on a couple Intel 545s in RAID1(that's how much I trust them), which I intended as a stopgap for Optane. But now I struggle to pick a suitable replacement, with grabbing some Samsung 970 Pros before they're gone seems like the best option.

Wherever I look in forums, YouTube etc. people seem to complain about SSDs getting worse. And while I get that the majority of buyers don't need super endurance and reliability, there should certainly be more than enough of us professionals and prosumers to justify at least one PRO SSD on the market. And when they put several hundred GBs of SLC cache in SSDs, why can't they make a pure SLC PCIe 4 SSD for the prosumer market? I'll gladly pay $200-250 per 512 GB, and I'm sure many others will too. :confused:

efikkanIn the places I've worked, people seem to "burn through" SSDs like crazy. I've seen many cases where machines get unstable due to file corruption, caused by defective SSDs(SMART errors), often after about 2-3 years of usage (in a few cases 1 year), and all of these machines have only been used for programming/compilation and light web browsing. Yet the few SSDs that did show statistics was way past their endurance rating, dozens of TB written, and it makes no sense because I know what the users have been doing. My theory is that somehow writing millions of tiny files causes the SSD to write larger cells plus the wear leveling going crazy, so 1 GB/day of files turns into "200 GB" worth of wear. I've been building computers since the 90s, and this just feels off to me, we seem to be approaching the reliability of floppies.

If my experiences are representative for people doing a lot of small writes quickly wearing out SSDs, then these endurance ratings are probably as misleading as the MTBF ratings on HDDs. I actually find it hard to find solid solutions when the storage mediums are so unreliable. And of course I have backups, that's not the problem, the problem is the time wasted on recovery and cleaning up etc. A couple of years ago I moved to a separate OS SSD on my desktops and my file server, and making the OS drive "disposable" (meaning it contains nothing important), if something goes bad I can format it (or replace it), reinstall the OS and run a few scripts and I'm back up in 15-30 min. In my home workstation, my workspace resides on a couple Intel 545s in RAID1(that's how much I trust them), which I intended as a stopgap for Optane. But now I struggle to pick a suitable replacement, with grabbing some Samsung 970 Pros before they're gone seems like the best option.

Wherever I look in forums, YouTube etc. people seem to complain about SSDs getting worse. And while I get that the majority of buyers don't need super endurance and reliability, it should certainly be more than enough of us professionals and prosumers to justify at least one PRO SSD on the market. And when they put several hundred GBs of SLC cache in SSDs, why can't they make a pure SLC PCIe 4 SSD for the prosumer market? I'll gladly pay $200-250 per 512 GB, and I'm sure many others will too. :confused:

My working theory is that SSDs have higher rates of bitrot than people expect.

Bitrot is a known problem on older drives: be they CDROMs or Hard Drives. In fact, its so well known, that these older mediums have amounts of resources dedicated to finding, and fixing, bitrot. So we computer users forget about that problem in practice.

SSDs are a bit newer. Like DRAM, SSDs work by "capturing electrons" on a gate. (Unlike DRAM, those electrons stay on the gate for days, weeks... months.). But do they stay on there for a year+ ??

I really don't think so. And typical filesystems (such as NTFS or EXT3) do NOT have bitrot protections. Only things like BTFS, ZFS (Linux/Unix) or ReFS (Windows), which are resilient filesystems that are reserved for enterprise users / higher-paying customers.

---------

The Samsung 830 EVO issues were practically a bitrot problem for example. If a file (or sector) was left unread for months, it'd be nearly impossible to read and the drive slows down dramatically trying to figure out what the hell is in that sector. That was back in the MLC days, before TBW was really an issue.

newtekie1Follow the chain of posts. Here I'll help you:

Tom Yum said: SLC cache doesn't help because game installs are too big.
I Responded to Ton Yum: But you're downloading those games from the internet, which is still slower than QLC.
Then you Responded to me: What does internet speed have to do with anything?

You decided to chime into a conversation about games and then want to act like we aren't talking specifically about games. Sorry, we were talking specifically about installing games. If you can't follow the conversation, don't chime in.

I NEVER mentioned games though. My reply was to your comment about internet speed. Sorry, but you're reaching here.

newtekie1Yeah, thats the oldest generation which isn't produced anymore. It was replaced by the 665p which has 200GB SLC cache for the 1TB model. As do most other QLC drives now.

If you'd read the Anandtech review I linked to, you would've known the 665p is the one that Intel has quietly discontinued and doesn't even mention any more, whereas, for whatever reason, they still sell the 660p.
So the baseline, even in Intel's own comparison of the 670p, is the 660p.

TheLostSwedeI NEVER mentioned games though. My reply was to your comment about internet speed. Sorry, but you're reaching here.

You replied to my comment which was itself a direct reply to a comment about games, I even quoted the post about games in the comment you replied to. Please at least attempt to keep up with the conversation if you're going to chime in.

TheLostSwedeIf you'd read the Anandtech review I linked to, you would've known the 665p is the one that Intel has quietly discontinued and doesn't even mention any more, whereas, for whatever reason, they still sell the 660p.
So the baseline, even in Intel's own comparison of the 670p, is the 660p.

Even still, is 140GB not enough SLC cache? Are you regularly transferring 100GB+ to your SSD? If so, then yeah, maybe QLC isn't for you. But that doesn't mean it won't work nicely as a system drive for everyone except you apparantly.

newtekie1Even still, is 140GB not enough SLC cache? Are you regularly transferring 100GB+ to your SSD? If so, then yeah, maybe QLC isn't for you. But that doesn't mean it won't work nicely as a system drive for everyone except you apparantly.

That's not how SLC cache works on modern systems. The SLC portion of the drive IS the TLC (or QLC) portion.

All you need to "lose" the SLC portion is to fill your drive to about 80% capacity. When the drive is 0% full, it uses all of the other bits as an SLC cache. But when those bits are filled with data, you've effectively "lost" the SLC portion and the drive begins to perform worse-and-worse.

That's why there are 80% full or 90% full tests. Historically, it may have been for TRIM reasons, but this SLC-cache thing today is yet another reason 80% full tests are useful.