Effect of SLC Caching on SSD Endurance

[blanarahul]

Kingston KC3000 Review - Faster Than Samsung 980 Pro

The Kingston KC3000 is built using the Phison E18 controller and Micron's best 176-layer TLC NAND flash. In our performance testing, the drive can beat the Samsung 980 Pro and is the fastest SSD we ever tested. It shares that performance throne with the WD Black SN850.

www.techpowerup.com

The Kingston KC3000 has 2000 GB of TLC cache. It can use almost all of it (1930 GB) in SLC mode for 1930/3 = 643 GB.

I keep wondering though. Isn't this writing to the NAND twice? Say I write 100 GB. First I consume 300 GB worth of NAND when writing in SLC mode. Then I consume 100 GB worth NAND in TLC mode. Of course, writing in SLC isn't nearly as harmful but it is somewhat harmful isn't it? We don't even get to choose if we are willing to let go of this SLC caching for better endurance.

@Chris_Ramseyer Can you offer some insights into how harmful (or harmless) SLC caching is to NAND endurance?

@W1zzard Can you offer some insights to this?

 

[Selaya]

writing to TLC in SLC mode causes about as much wear as writing to SLC in SLC mode - compared to actual TLC writes like, next to none at all.

 

[Mussels]

SLC caching uses less writes, you could think of it like anything written in SLC mode uses 1/4 the lifespan of QLC writes


my numbers there is made up, we'd need an expert to tell us how much it actually helps. Short version is, keep some free space on the drive if you want it to live longer.

 

[blanarahul]

My numbers there is made up, we'd need an expert to tell us how much it actually helps. Short version is, keep some free space on the drive if you want it to live longer.

I read online that a 2D NAND cell in TLC mode has a life of 3k writes, in MLC has a life of 10k writes and in SLC mode has a life of 100k writes. Based on this if you do the math, the net effect of writing to the NAND chip first in SLC mode and then in TLC mode is about 10%. Meaning what would have been a life of 3000 cycles becomes closer to 2750 cycles. (See the math below)

However, not only is all this based on the assumption that a TLC chip in SLC mode will have 30x the endurance. I am also (incorrectly) assuming identical failure point for both. A NAND chip which has reached the point of failure in TLC mode should still be perfectly okay to be used in SLC mode. This means that SLC mode cannot have 30 times the write cycles. It might be closer to 10x or 20x, I don't know. Which is exactly why I asked the question. How much endurance are we sacrificing for increased write performance? Large read performance will become useful for gamers because of DirectStorage in 5-6 years so having that as an option might become necessary.

But how much write performance do most consumers need? Aren't their needs ultimately limited by their internet bandwidth? I am pretty sure most people don't have a 4 Gbps/500 MBps internet connection. Some might have a 5 Gbps/600 MBps external SSD but writing directly to TLC can already accomplish that.

Intel's 760p had a sequential write speed of 560 MBps and when compared to a similar drive (ADATA SX8200) which had 1660 MBps sequential writes, but identical read performance... the Adata was barely 6% faster overall in TPU's review because the 2 drives had basically identical real life performance for most uses (Photoshop Editing and ISO File Copy being the biggest differentiators).

	P/E Cycles	Capacity in GB	Total Life in GB	%age life used for 10 GB write
SLC	100,000​	10​	1,000,000​	0.001%​
TLC	3,000​	30​	90,000​	0.011%​
			Total %age life used	0.012%​
			Effective Life in GB	82,569​
			Effective P/E	2,752​

 

[Mussels]

No you dont lose any endurance from SLC mode, you gain it.

SLC mode *reduces* the number of write operations, since less cells have to write.

 

[blanarahul]

No you dont lose any endurance from SLC mode, you gain it.

SLC mode *reduces* the number of write operations, since less cells have to write.

Allow me to explain how NAND works.

The smalled unit of storage in a NAND chip is called a page. The size of this would be 16 KB in TLC Mode or 5.33 KB in SLC Mode. This is the smallest amount you are allowed to write/read. Whether you are writing/reading 1 KB to your SSD or 16 KB, the SSD will have to read/write that whole page.

64 to 512 of these pages make a block.

Hundreds of these blocks are then organized into slices.

Each NAND chip is made of up multiple slices.

The parallelism in SSDs comes from having multiple slices and multiple chips.

Say, your SSD has 4 NAND chips each of which has 4 slices. That is 16 slices total. Your SSD controller has 16 connections to these 4 NAND chips. One for each slice.

You want to write 80 KB to your SSD.

Step 1 - break that 80 KB into fifteen 5.33 KB chunks

Step 2 - each 5.33 KB is written to a SLC mode page in each slice. this is done for all 15 chunks to 15 pages in 15 slices, parallely/together.

Btw, it is because of this parallel reading and writing that SSDs get faster (MBps) as your file gets larger.

Step 3 - The 80 KB worth of data that was written in SLC mode is written again to five 16 KB TLC mode pages (distributed across 5 slices).

Step 4 - Those fifteen 5.33 KB SLC pages are then emptied to be used again in the future.

Steps 3 and 4 taken together are called "flushing the SLC cache to TLC".

This is the difference I am talking about.

Without SLC caching, I would have just consumed 5 TLC pages.

With SLC caching, I am consuming 15 SLC pages in addition to the 5 TLC pages.

This is why SLC caching reduces SSD endurance by some (as of yet undetermined) amount.

 

[AlLOL2001]

I've the same question, I'm researching before buying a SSD, I think that I'll go with the Samsung 980...

My new question is: there is someway to disable the SLC cache?

I'm not sure how do the "Intelligent TurboWrite" works, but I really hope that there's someway... If somebody know how to disable it, please tell me...

I've read something like Intel can disable that SLC cache...

Sorry, english isn't my first language... TY in advance...

 

[joemama]

I've the same question, I'm researching before buying a SSD, I think that I'll go with the Samsung 980...

My new question is: there is someway to disable the SLC cache?

I'm not sure how do the "Intelligent TurboWrite" works, but I really hope that there's someway... If somebody know how to disable it, please tell me...

I've read something like Intel can disable that SLC cache...

Sorry, english isn't my first language... TY in advance...

I don't think the SLC cache function is something you can control from the computer side since it is written in the SSD controller

 

[chrcoluk]

It depends how intelligent the controller is.

A SSD insider revealed sometime in 2020 that there is primitive controllers that write almost everything that's written to SLC cache again to TLC, then back to SLC again if the file is updated, so whilst SLC cache writes might be good for endurance, it doesn't mean much if they still have to be written again on TLC.

He developed a drive working with AMD, that had more intelligent behaviour files that are mostly writes would stay in SLC all the time e.g. logs, for this to work files would generally stay on SLC for a while whilst the drive figures out if they read or write heavy, and only after this delay read heavy files get moved to TLC, and with this new behaviour the expected endurance of the drive far exceeded 3D MLC.

Many existing drives which just go for headline benchmark results, I believe start moving data from SLC quite quickly to try and maximise performance for new writes.

 

[bug]

writing to TLC in SLC mode causes about as much wear as writing to SLC in SLC mode - compared to actual TLC writes like, next to none at all.

I doubt that very much. Whether you write 0 and 1 or 0..7, you're writing to the same physical cell. You're using one p/e cycle still.

@blanarahul 3D TLC brings the p/e cycles in line with planar MLC. Endurance of these drives is more than enough. For example my 850EVO is over 5 years old, has been used a gaming and then as an OS drive and still displays over 90% integrity.

 

[blanarahul]

Many existing drives which just go for headline benchmark results, I believe start moving data from SLC quite quickly to try and maximise performance for new writes.

I believe this to be the case as well.

Back in the day we could choose between having TLC without SLC caching and TLC with SLC caching. Nowdays we don't get that choice anymore.

This is what Samsung's TLC without SLC caching SSD looks like. In comparison, with SLC caching (980 Pro) -

Random write performance sees quite the jump from 100k IOPS to 1000k IOPS and Seq Write numbers are 2.4 GBps for a 2 TB class drive. So it would definitely look quite bad for marketing to not have such performance. But if you ask me, having a 1 TB SSD with 1.2 GBps sequential writes would be more than okay.

---

---

---

This is what Kingston says about their TLC without caching drive.

Had they used a PCIe 4.0 controller they would be able to reach 7000 MB/sec reads no problem. But that 925 MB/sec will be hard to advertise. On the other hand, endurance numbers are a lot better for this drive - 1095 TB for their 960 GB model (compared to 600 TB for consumer 1 TB ssds).

 

[GabrielLP14]

Very interesting content, however you guys failed to mention one details that's a drawback, WAF. Write Amplification Factor

 

[blanarahul]

WAF. Write Amplification Factor

How is that relevant to our discussion? Can you please explain?

 

[Tetras]

I believe this to be the case as well.

Back in the day we could choose between having TLC without SLC caching and TLC with SLC caching. Nowdays we don't get that choice anymore.

This is what Samsung's TLC without SLC caching SSD looks like. In comparison, with SLC caching (980 Pro) -

Random write performance sees quite the jump from 100k IOPS to 1000k IOPS and Seq Write numbers are 2.4 GBps for a 2 TB class drive. So it would definitely look quite bad for marketing to not have such performance. But if you ask me, having a 1 TB SSD with 1.2 GBps sequential writes would be more than okay.

---

---

---

This is what Kingston says about their TLC without caching drive.


Had they used a PCIe 4.0 controller they would be able to reach 7000 MB/sec reads no problem. But that 925 MB/sec will be hard to advertise. On the other hand, endurance numbers are a lot better for this drive - 1095 TB for their 960 GB model (compared to 600 TB for consumer 1 TB ssds).

I don't think the ironwolf 110 or 125 pro have a slc cache either and also much higher endurance.

 

[Black [Super Saiyan Rosé]]

How is that relevant to our discussion? Can you please explain?

Dynamic SLC design lead to a worse WAF (data may first be written as SLC and later be rewritten as TLC/QLC), while static one no. In fact, a SSD with dynamic SLC cache doesn't improve endurance, while an other one with static SLC yes, static SLC can improve NAND endurance to like 30-40K PEC (P/E cycles) - it depends on the NAND model.

 

[bug]

Dynamic SLC design lead to a worse WAF (data may first be written as SLC and later be rewritten as TLC/QLC), while static one no. In fact, a SSD with dynamic SLC cache doesn't improve endurance, while an other one with static SLC yes, static SLC can improve NAND endurance to like 30-40K PEC (P/E cycles) - it depends on the NAND model.

Whether the cache is static or dynamic, it's still just a cache (i.e. data gets written to it and then flushed to the main storage). What's the difference here? What am I missing?

 

[Selaya]

dynamic slc cache contents get written into tlc at some point (either when the drive fills or when it's idle, depending on the configuration of the firmware), which means that data gets written twice (into slc cache, then into tlc proper) = write amplification

 

[Black [Super Saiyan Rosé]]

Whether the cache is static or dynamic, it's still just a cache (i.e. data gets written to it and then flushed to the main storage). What's the difference here? What am I missing?

The fact that dynamic and static - even if we are talking about cache - aren't the same thing, they work in different way and have different advantages/disadvantages. The static SLC design is in the OP (over-provisioning) space, is always available and always in SLC mode for the device's life, the dynamic SLC design depends on "how much space the user is using, diminishing in size as the drive is filled.".

 

[R0H1T]

Steps 3 and 4 taken together are called "flushing the SLC cache to TLC".

Then you should also know that not all the data from SLC (or MLC cache) is flushed to the disk! Only the final writes, & that too if necessary!

Deferred caching, with PrimoCache is the easiest way to learn this ~ only final writes, or urgent writes when cache is full, is generally flushed to the disk! This saves a lot of unnecessary writes & is actually a major way to save on your disk's lifetime endurance. Whatever napkin math you're doing doesn't translate to real world numbers, there's literally a ton of variables involved.

 

[blanarahul]

Dynamic SLC design lead to a worse WAF (data may first be written as SLC and later be rewritten as TLC/QLC)

That's exactly what I said in my original post without calling it something fancy like WAF.

static SLC can improve NAND endurance to like 30-40K PEC

By how much? Because whether you write to TLC directly or funnelled through SLC cache, you are writing to TLC either way. How is it improving TLC's endurance?

 

[bug]

The fact that dynamic and static - even if we are talking about cache - aren't the same thing, they work in different way and have different advantages/disadvantages. The static SLC design is in the OP (over-provisioning) space, is always available and always in SLC mode for the device's life, the dynamic SLC design depends on "how much space the user is using, diminishing in size as the drive is filled.".

Still, each write means one write to the cache and another one to the main storage. No difference in write count.

 

[Black [Super Saiyan Rosé]]

I don't think the ironwolf 110 or 125 pro have a slc cache either and also much higher endurance.

From Real Hardware Reviews review (IronWolf 110): "At its heart this is a pseudo-SLC cache buffer; however, unlike most TLC solid-state drives which use a fixed capacity for their pseudo-SLC cache, DuraWrite is a complete floating pseudo-SLC cache buffer that will use every bit’s worth of free space on the drive."

 

[blanarahul]

Then you should also know that not all the data from SLC (or MLC cache) is flushed to the disk! Only the final writes, & that too if necessary!

I highly doubt this. What is your source on this?

 

[R0H1T]

You mean apart from some/many of the SSD reviews?

It's not 100% copy of PrimoCache because SLC caches come in lots of different sizes & are implemented differently. But the net result is mostly the same.

 

[Black [Super Saiyan Rosé]]

By how much? Because whether you write to TLC directly or funnelled through SLC cache, you are writing to TLC either way.

Usually TLC flash is around 2000-3000 PEC, but also here depends on the model, because, for example, Micron B37R (128L) is 5000 PEC, while SpecTek one (which is a sub-brand with worse bin) is like 700.

How is it improving TLC's endurance?

I think for the fact that static SLC is in OP or other non-accessible SSD's spaces, not like dynamic one which its size depends on the space in the SSD and other factors.

Still, each write means one write to the cache and another one to the main storage. No difference in write count.

Maybe yes, but in terms of SSD'/NAND's endurance no.