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

[Rifter]

I'd say a regular 860EVO or Intel 545s should handle that without issues. As cbn points out, 4TB a month isn't really that much in the larger scheme of things*. I'd also see if I couldn't get a 2TB drive for the additional physical NAND, that way you'll only use 2 P/E cycles a month, since there are more physical cells present.

If you want additional security, there is really only the 860PRO available. It has a 1200TBW rating for the 1TB model (which is competitive with low-end enterprise drives), and uses MLC NAND. So should be able to withstand some abuse. But it is more expensive in the larger capacities.

That is unless you're willing/able to consider enterprise class gear.

*The spectre of write amplification might be an issue, but with mostly sequential writes shouldn't.

Yeah sounds like the 860 pro would be ideal, but the EVO at half the price will get the job done. Thanks.

 

[cbn]

According the following news post the BX500 is shipping with 96L TLC for the 960GB version:

https://www.anandtech.com/show/14039/micron-announces-1300-client-sata-ssd-with-96l-tlc

So I looked up a review for the drive and found the following:

https://www.tweaktown.com/reviews/8...gb-ssd-review-micron-96l-tlc-ships/index.html

One interesting point is that Crucial is using its Momentum cache software to buffer writes:

Crucial's Momentum Cache software will help to decrease the number of random writes to the drive by buffering the small blocks in DRAM and passing the data to the drive as large block data. This also helps with to decrease the size of the flash transition table map.

EDIT: Some more info on Crucial Momentum cache here.

Additionally, Momentum Cache reduces wear on an SSD by filtering repetitive data written to the drive.

(The following, from the same link, explains how much System DRAM is used by Momentum cache)

When enabled, Momentum Cache is optimally set at one-quarter the size of the physi-cal system memory, up to a maximum of 4GB. This memory allocation dynamically scales in direct relation to the amount of memory consumed by applications residing on the system.

As application memory use increases, Momentum Cache relinquishes memory to en-sure the applications are not impacted. When the amount of idle memory available tothe system falls below 1GB, Momentum Cache reduces its memory consumption to maintain at least 1GB of idle memory available to the system. Memory allocation be-tween the system and Momentum Cache is continually evaluated to maintain this bal-ance.

During intense application memory usage, Momentum Cache reduces its memory allo-cation to a minimum value of 256MB. If additional memory resources are required, Mo-mentum Cache relinquishes all memory to the system. In this scenario, all data passes between the system and the SSD untouched by Momentum Cache.

Momentum Cache does not re-allocate idle memory until a minimum of 1.4GB of sys-tem memory becomes available. Momentum Cache continues to increase its allocationof memory as it becomes idle in the system.

P.S. TBW for the 960GB BX500 is 240.

 

[cbn]

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.

I wonder if we will end up seeing any small die 3D QLC drives that regain some of that lost write?

With that mentioned, while preparing a comparison of 860 QVO vs. 660p vs. BX500 I was very surprised to see both the 480GB (64L 3D TLC) and 960GB (96L 3D TLC) BX500 have relatively slow sustained writes:

I wonder why that is?

(I was expecting sustained write to be BX500 strong suit when compared to 860 QVO and 660p, but as you can see it is actually quite slow.)

 

[cbn]

Here is the sustained Sequential write from another SM2259XT/IMFT 3D TLC SSD (Mushkin Source):

https://www.anandtech.com/show/13421/the-mushkin-source-sata-ssd-review/5

So just like the BX500 it is much slower than MX500 which also uses IMFT 64L 3D TLC, but with SM2258 rather than SM2258XT. (Intel 545s also uses IMFT 64L 3D TLC ( specifically the small 256Gb die), but with SM2259 rather than SM2258).

With that noted, how much of this difference in Sustained Sequential (vs. MX500) is due lack of DRAM buffer?

 

[cbn]

According to Anandtech the 1TB Crucial MX500 uses the 256Gb 64L IMFT 3D TLC die (ie, the small die):

https://www.anandtech.com/show/12165/the-crucial-mx500-1tb-ssd-review

And comparing the second chart in post #78 to the second chart in post #79 we see that the 1TB MX500 has ~double (~400 MB/s) the sustained Sequential write of the 207 MB/s 500GB MX500. That makes sense because the 1TB MX500 has double the parallelism of the 500 MX500.

But what to make of the ~70ish GB/s write of 500GB Mushkin Source and 480GB Crucial BX500?*

Are these dram-less SM2258XT SSDs using 64L IMFT 768Gb 3D TLC dies? (ie, the largest die)

It does look like that is the case.

*The 960GB Crucial BX500 uses 96L IMFT 3D TLC that actually looks like it has even less parallelism per GB than (what I suspect is) the large die 64L 3D TLC (reason: ~75 MB/s for 960GB is 50 % less Sequential write than I would expected for that capacity if it were using the same 64L 3D TLC dies as the 480GB BX500).

 

[cbn]

As I mentioned here I do wonder if IMFT 96L (as used in 960GB BX6500) is based on a lithography smaller than 20nm.

At the same time I don't think IMFT would be afraid of going with a large 230mm2 die because Intel has specialized products like Ruler form factor SSDs.

However, it is still interesting to think about 16nm being used. This especially as both Intel and Micron produce 3DXpoint (which could be used to buffer small writes to the NAND)