Question Are dramless TLC and QLC SSDs best to be avoided?

[whm1974]

After reading several reviews of both dramless TLC and QLC SSDs over about a month i have come to the conclusion that such SSDs are basically worthless and and to be avoided like a rapid dog. Decent TLC w/ DRAM caches SSDs are not that much more and the offer better performance and longevity.

It starting to look NAND Flash has begin to reach practical limits to how many bits per cell can be had.

 

[UsandThem]

They have their place as even the slowest SSDs are faster than a spinner.

However, right now QLC doesn't make a whole lot of sense because the pricing is too high (like the new Samsung QLC drive). But with SSD pricing falling so much over the last year, it's only a matter of time before the entry level products like QLC drive fall below the mid-range products. Right now they are simply priced out of their intended market.

 

[whm1974]

You might be right about them having their place once prices drop. But I think I would stick with TLC SSDs for the foreseeable future.

 

[VirtualLarry]

They have their place as even the slowest SSDs are faster than a spinner.

Not so sure about that. Samsung QVO drives, hit 80MS/s writes, once they run out of SLC write cache. That's half the sequential write speed of a decent spinner.

 

[whm1974]

Not so sure about that. Samsung QVO drives, hit 80MS/s writes, once they run out of SLC write cache. That's half the sequential write speed of a decent spinner.

In which case I'll pass on QLC SSDs. Unless it for something I won't be writing to very often.

 

[UsandThem]

Not so sure about that. Samsung QVO drives, hit 80MS/s writes, once they run out of SLC write cache. That's half the sequential write speed of a decent spinner.

Which only advanced or power users would hit in most scenarios. But Joe Smo who has one in a general use PC will not hit, and they still will have the benefit of QLC drives being faster than a HDD as an OS drive.

PC enthusiasts like us, or power users who write or transfer large files know not to buy them.

 

[IntelUser2000]

But Joe Smo who has one in a general use PC will not hit, and they still will have the benefit of QLC drives being faster than a HDD as an OS drive.

Ooh, I don't know about this. I had a cheap DRAM-less drive. It became slow within a week. Basic browsing and installing became noticeably slower. I really did not do anything remotely intensive on it. After TRIM it became significantly faster. After a week of usage.

 

[whm1974]

Ooh, I don't know about this. I had a cheap DRAM-less drive. It became slow within a week. Basic browsing and installing became noticeably slower. I really did not do anything remotely intensive on it. After TRIM it became significantly faster. After a week of usage.

Did you replace it with a proper one that has cache?

 

[UsandThem]

Ooh, I don't know about this. I had a cheap DRAM-less drive. It became slow within a week. Basic browsing and installing became noticeably slower. I really did not do anything remotely intensive on it. After TRIM it became significantly faster. After a week of usage.

I don't doubt it, but there are different drives, with different strengths and weaknesses. Some of those DRAMless drives are junk. BTW, what drive did you buy that was that bad?

But as far as the available QVL drives go, several are PCIe, and offer pretty good performance compared to SATA SSDs, albeit at a higher price.

The Samsung QVL SATA SSD, as shown in the Anandtech review, is faster than a HDD in almost every test. There are a few scenarios where it's not much faster, but overall they are faster than a spinner.

Now personally, I'd spend the few extra bucks on a good drive, but once these larger QVL SSDs get cheaper, they would be a decent mainstream choice IMO.

 

[cbn]

One thing the DRAM buffer does (with help of the controller) is reduce write amplification:

https://www.tomshardware.com/reviews/Intel-x25-m-SSD,2012-5.html

"Cache Needed for Write Amplification Control

This is why Intel’s controller needs the DRAM buffer. It is used as a short term memory to store data, so it can execute write operations more efficiently than a conventional Flash controller, which typically triggers writes one by one, writing huge cells even if the smallest data chunks need to be stored."


With that noted, I wonder if write amplification would be hypothetically worse on a DRAM-less 3D QLC drive compared to a DRAM-less 3D TLC drive?

Is the Block size larger on a 3D QLC drive compared to a 3D TLC drive? If it is then that would mean a higher amount of write amplification all things being equal.

 

[IntelUser2000]

Did you replace it with a proper one that has cache?

No, because its a secondary system anyway. Besides, its not totally useless.

Silicon Power S60 slim. If you search for it, most reviews are about a different version. The manufacturer page is different, and unlike most reviews doesn't have any DRAM cache. There's like one review that gave it a 5-star rating after testing it once with CrystalDiskMark.

Newegg reviews on it were pretty decent. Reviews can be different from how it actually is. I bet most such caching can be overwhelmed rather easily. That's why TRIM exists.

 

[cbn]

Silicon Power S60 slim. If you search for it, most reviews are about a different version. The manufacturer page is different, and unlike most reviews doesn't have any DRAM cache.

Yep, there two versions:

1st version: SandForce SF-2281 which is indeed DRAM-less and this paired with Intel NAND --> https://www.tweaktown.com/reviews/6868/silicon-power-s60-240gb-ssd-review/index.html

2nd version: Phison S8 (has dram buffer) and 15nm Toshiba TLC --> https://www.tweaktown.com/reviews/6868/silicon-power-s60-240gb-ssd-review/index2.html

But I believe the first version uses MLC NAND because I didn't see support for TLC in the link I provided for SF-2281 and also the following S60 Slim SSD has an option for MLC--> https://www.newegg.com/Product/Product.aspx?Item=0D9-0021-00006 (EDIT: Silicon Power Website on the following webpage (under features) also says the drive also uses MLC NAND---> https://www.silicon-power.com/web/product-Slim_S60)

"Adopts MLC NAND flash and "SLC Cache Technology" to improve overall performance"

 

[IntelUser2000]

I have the purple colored one.

"Adopts MLC NAND flash and "SLC Cache Technology"

That's why it sucks. SLC is better than MLC but doesn't hold a candle to DRAM. SLC NAND needs controller trickery so it doesn't wear out and for optimal performance, just not much as MLC. Therefore, its going to go through the same degradation in performance after its dirty. Now that makes it a bad cache.

Sure in benchmarks it does well. In every real world scenario, my 9 year old X25-M with 250MB read/70MB write still kicks its ass. It did pay to pay back then.

 

[hojnikb]

Thats why TRIM exits, so SSDs dont get "dirty". Last time i checked, x25m does not support it

 

[IntelUser2000]

Thats why TRIM exits, so SSDs dont get "dirty". Last time i checked, x25m does not support it

The first gen(which I have) doesn't. Second gen does.

X25-M doesn't need it as badly because even on its dirty state the system doesn't feel as slow as the S60. S60 NEEDS TRIM to function.

 

[Charlie22911]

It depends.
SATA SSDs without DRAM for mapping functions would likely (depending on design) face reduced endurance and performance.
NVMe SSDs without DRAM on the other hand could (again, depending on design) use host memory for its mapping table.

I don’t have experience with either of the above implementations but intuition tells me that in the latter case performance impact would be measurable due to increased latency of reaching out to system ram, but imperceptible in common workloads if implemented well.

 

[cbn]

According to Anandtech 2018 SSD in review 3D QLC has (@ .1 drive writes per day for 5 years*) about 1/3 the write endurance as low cost 3D TLC.

That is consistent with previous reports I have seen.

 

[cbn]

According to this report Intel is using 20nm lithography for their 64L 3D NAND, but I wonder how soon they would use the 16nm half node?

If so, how would that affect their endurance?

Or do they expect a good decrease in endurance, but plan on countering that by using Optane in hybrid fashion? (See link and quote below for an example of what I am referring to)

https://www.tomshardware.com/news/intel-optane-memory-qlc-cache,37223.html

Sitting behind high-endurance Optane Memory (cache) will remove any immediate concerns regarding QLC's endurance. A bundle that gives users a free or very low cost 512GB boot drive will be difficult to pass over even for the most adamant Optane and QLC naysayers.

P.S. Would be very interesting to see 2048Gb 96L 3D QLC from Intel. A nice side benefit would be that this type of high capacity 3D QLC die would be a good fit for Maxio technology MAS0902 dram-less SATA controller (which unlike the dram-less Phison S11 controller has a much higher capacity( 2TB (or 4TB) for MAS0902 vs. only 1TB for S11)).

 

[cbn]

According to the Anandtech post on ADATA SU630 Ultimate (DRAM-less SATA with 3D QLC) the drive has 50TB TBW for 240GB capacity and 100TB and 200TB TBW for 480GB and 960GB capacities respectively.

This works out to be 208 P/E cycles. (re: 50TB/.240TB = 208, 100TB/.480TB = 208, 200TB/.960TB = 208).

That doesn't sound like much when we compare that figure to what is seen in the chart below:

Furthermore, ADATA is only offering a 2 year warranty on this drive.

EDIT (Correction to what I wrote above): TBW and P/E are calculated differently than what I thought --> https://www.anandtech.com/show/7947/micron-m500-dc-480gb-800gb-review/3

 

[VirtualLarry]

Maybe in the near / short-term future, what with SATA and PCI-E 3.0 x4 interfaces nearly saturated with today's controllers, that, barring unforseen random failure events, perhaps we should buy SSDs (within a performance class) solely by TBW / $?

IOW, buy the highest-endurance SSD within a performance class, for the lowest $ per TBW.

This is something that I would like to see as a metric in future SSD reviews, both here at AT, and at other sites like TPU. (With those wonderful charts.)

 

[cbn]

Here are TBW for Crucial BX500 (DRAM-less SATA using 3D TLC):

https://www.tomshardware.com/reviews/crucial-bx500-ssd,5377.html

120GB: 40TB
240GB: 80TB
480GB: 120TB

(So not that much better than the ADATA SU630 Ultimate which has 50TB for 240GB and 100GB for 480GB))

Interesting that the 480 GB doesn't have 160GB TBW when the 120GB and 240GB are 40GB and 80GB respectively. Maybe Crucial uses lower bin 3D TLC for the 480GB since it has more capacity?

P.S. BX500 uses SM2258XT dram-less controller and Micron 64L 3D TLC NAND.

 

[cbn]

Looking back at the Anandtech Crucial BX200 480GB review both the 480GB BX200 (SM2256 with micron 16nm TLC) and 500GB BX100 (SM2246EN with micron 16nm MLC) are listed at 72TB for TBW.

(So in these two cases SATA with dram-buffer and either 16nm TLC or 16nm MLC have less endurance than the SATA DRAM-less 3D QLC SSD mentioned in post #19).

P.S. I am guessing the reason the 480GB BX200 does as well as the 500GB BX100 in TBW has to do with the greater amount of error correction on the SM2256 controller compared to the SM2246EN controller.

 

[cbn]

Maybe in the near / short-term future, what with SATA and PCI-E 3.0 x4 interfaces nearly saturated with today's controllers, that, barring unforseen random failure events, perhaps we should buy SSDs (within a performance class) solely by TBW / $?

IOW, buy the highest-endurance SSD within a performance class, for the lowest $ per TBW.

For SATA the highest endurance SSDs I have found so far are the Intel 545s and the Samsung 860 EVO.

512GB Intel 545s has 288TB TBW and the 500GB Samsung 860 EVO has 300TB TBW.

And these probably have the best TBW/$ ratio as well.

 

[Glaring_Mistake]

According to the Anandtech post on ADATA SU630 Ultimate (DRAM-less SATA with 3D QLC) the drive has 50TB TBW for 240GB capacity and 100TB and 200TB TBW for 480GB and 960GB capacities respectively.

This works out to be 208 P/E cycles. (re: 50TB/.240TB = 208, 100TB/.480TB = 208, 200TB/.960TB = 208).

That doesn't sound like much when we compare that figure to what is seen in the chart below:

Furthermore, ADATA is only offering a 2 year warranty on this drive.

TBW can be pretty loosely based on endurance though.

Here are TBW for Crucial BX500 (DRAM-less SATA using 3D TLC):

https://www.tomshardware.com/reviews/crucial-bx500-ssd,5377.html

120GB: 40TB
240GB: 80TB
480GB: 120TB

(So not that much better than the ADATA SU630 Ultimate which has 50TB for 240GB and 100GB for 480GB))

Interesting that the 480 GB doesn't have 160GB TBW when the 120GB and 240GB are 40GB and 80GB respectively. Maybe Crucial uses lower bin 3D TLC for the 480GB since it has more capacity?

P.S. BX500 uses SM2258XT dram-less controller and Micron 64L 3D TLC NAND.

Probably just that they think a 120TBW rating is good enough for a budget drive like BX500.
For example the WD Blue 3D, TBW is 100TB at 250GB, 200TB at 500GB, 400TB at 1TB and then just 500TB at 2TB.
Or MX300 where the 1TB capacity has a TBW rating of 360TB but the 2TB capacity has a TBW rating of just 400TB in comparison.
Think that the BX500 may use NAND of a bit lower quality than the MX500 though like they did with the MX100/BX100 (former rated for 3000 P/E, latter for 2000 P/E).

Looking back at the Anandtech Crucial BX200 480GB review both the 480GB BX200 (SM2256 with micron 16nm TLC) and 500GB BX100 (SM2246EN with micron 16nm MLC) are listed at 72TB for TBW.

(So in these two cases SATA with dram-buffer and either 16nm TLC or 16nm MLC have less endurance than the SATA DRAM-less 3D QLC SSD mentioned in post #19).

P.S. I am guessing the reason the 480GB BX200 does as well as the 500GB BX100 in TBW has to do with the greater amount of error correction on the SM2256 controller compared to the SM2246EN controller.

Not quite.

The BX100 and BX200 both have a TBW rating of 72TB at the 480-500GB capacities but that's because they have a 72TB TBW rating for all capacities (starting at 120GB for the BX100 and 240GB for the BX200).
A TBW rating of 72TB at 120GB and 240GB respectively implies that the BX100 should have double the endurance of the BX200 which since the BX100 is rated for 2000 P/E and the BX200 1000 P/E sounds fairly accurate.



More general on TBW:
The 860 EVO has (especially for the larger capacities) significantly higher TBW than the 850 EVO did despite still being specced for 2000 P/E.
Though if you doubled the TBW for the 850 EVO then you'd get the same TBW as for the 860 EVO (850 EVO may start at 75TB but that is at 120GB).
Even more interesting though is that while the TBW for the 860 Pro has gone up it looks to be specced at about 2000 P/E as well (in comparison the 850 Pro was specced for around 6000 P/E).

 

[Insert_Nickname]

For SATA the highest endurance SSDs I have found so far are the Intel 545s and the Samsung 860 EVO.

512GB Intel 545s has 288TB TBW and the 500GB Samsung 860 EVO has 300TB TBW.

And these probably have the best TBW/$ ratio as well.

Just to add a bit of perspective to the endurance of SSDs, my most abused SSD (Intel 600p 512GB/288TBW/32L 1st gen Intel/Micron NAND) has "only" clocked 6.6TB worth of writes in 3 years. It certainly hasn't been spared, so I guess this is worse then what your average consumer will ever do to it.

It'll be long obsolete before it hits 288TBW. At this rate, it'll last 127 more years before hitting the write limit...

In other words, I don't think endurance will ever be a problem for consumer drives.