Help me understand NAND parallelism as it relates to low end SSD controllers

[cbn]

In my ITT: We predict how low the price floor on 120GB/128GB SSD drops thread it was brought up that future larger capacity dies could still maintain the same parallelism as a lower capacity NAND die provided enough planes were present to compensate (ie, 256 Gb V-NAND die with 2 planes = same parellelism as 128 Gb V-NAND die with one plane).

So I got to wondering as these NAND dies increase in capacity to 512 Gb and 1024 Gb how many planes would be necessary to saturate a low end 120GB/128GB SSD with two channel SSD controller as far as sequential reads go?

 

[Hellhammer]

Sequential reads are never going to be an issue because reading from NAND is so much faster than writing to it (over 10x difference). Even a single 1024Gbit two-plane die should be able to saturate SATA 6Gbps interface. Write performance is the real issue because it's heavily impacted by NAND parallelism, although SLC caching is a good workaround to boost write performance even with a very small number of dies.

However, it's quite unlikely that we'll see 128GB SSDs with 512Gbit or 1024Gbit dies. As I mentioned in the thread you linked, 128GB SSD prices are already reaching their bottom and will be going the way of 32GB and 64GB SSDs this year (i.e. new drives are 256GB and up). Once 512Gbit dies are shipping 256GB is going to be the new smallest capacity.

 

[cbn]

Sequential reads are never going to be an issue because reading from NAND is so much faster than writing to it (over 10x difference). Even a single 1024Gbit two-plane die should be able to saturate SATA 6Gbps interface. Write performance is the real issue because it's heavily impacted by NAND parallelism, although SLC caching is a good workaround to boost write performance even with a very small number of dies.

However, it's quite unlikely that we'll see 128GB SSDs with 512Gbit or 1024Gbit dies. As I mentioned in the thread you linked, 128GB SSD prices are already reaching their bottom and will be going the way of 32GB and 64GB SSDs this year (i.e. new drives are 256GB and up). Once 512Gbit dies are shipping 256GB is going to be the new smallest capacity.

I looked up the specs on the 120GB 850 EVO (which uses 32 layer 128 Gb TLC V-NAND which has the same parallelism as 48 Layer 256 Gb TLC V-NAND) and was quite surprised to see the high published sequential writes of 520 MB/s (This confirmed with real world testing as reported by Newegg reviewers)

With two 512 Gb dies (each with two planes) having half the parallelism of the 120 GB 850 EVO would that come out to be 260 MB/s for sequential write? If so, then that would still be faster than SSD models like the 120 GB BX100 which uses SM2246EN controller and MLC NAND for an advertised sequential write spec of 185 MB/s.

So maybe Samsung doesn't produce any drives below 256 GB with 512 Gb dies, but I am actually wondering if other SSD vendors would?

 

[Hellhammer]

I looked up the specs on the 120GB 850 EVO (which uses 32 layer 128 Gb TLC V-NAND which has the same parallelism as 48 Layer 256 Gb TLC V-NAND) and was quite surprised to see the high published sequential writes of 520 MB/s (This confirmed with real world testing as reported by Newegg reviewers)

With two 512 Gb dies (each with two planes) having half the parallelism of the 120 GB 850 EVO would that come out to be 260 MB/s for sequential write? If so, then that would still be faster than SSD models like the 120 GB BX100 which uses SM2246EN controller and MLC NAND for an advertised sequential write spec of 185 MB/s.

So maybe Samsung doesn't produce any drives below 256 GB with 512 Gb dies, but I am actually wondering if other SSD vendors would?

520MB/s is with TurboWrite (i.e. SLC caching). Once the TurboWrite buffer is full (3GB in 120GB 850 EVO), the sequential write speed drops to 150MB/s. We've been transparent about that (link 1, link 2) and obviously most transfers in client workloads never touch the TLC portion until completion as writing >3GB of data at once is quite rare.

With a two-plane 512Gbit die the TLC write performance would likely drop below 100MB/s as parallelism would be cut in half and improved programming algorithms can't fully offset the difference. However, SLC cache performance would likely still be close to 500MB/s.

It's of course possible that we'll see 128GB drives with 512Gbit dies, but the write performance would be quite abysmal, especially if there's no SLC caching employed. I somewhat doubt that because by the time 512Gbit dies are shipping, 256GB SSDs will be nearly the same price as 128GB drives are today, and there's just no more room for 128GB drives to get cheaper.

Note that I'm not speaking officially here, so everything I've said is just my own speculation and analysis. Obviously I wouldn't be allowed to provide any specifics of our unreleased products, such as the V-NAND roadmap or NAND configuration.

 

[cbn]

From the Anandtech ADATA SP550 review:

http://www.anandtech.com/show/10131/the-adata-premier-sp550-ssd-review/7

850 EVO 120GB scores 166.93 MB/s, So factoring in parallelism being halved on the future Samsung 512 Gb TLC V-NAND a 120GB SSD based on two of these dies should come out around 83 MB/s. This assuming the same controller was used.

In contrast, 120GB ADATA SP550 SSD (SM2256 controller with SK Hynix 16nm planar TLC NAND) is only getting 65.05 MB/s on the sequential write.

 

[VirtualLarry]

520MB/s is with TurboWrite (i.e. SLC caching). Once the TurboWrite buffer is full (3GB in 120GB 850 EVO), the sequential write speed drops to 150MB/s. We've been transparent about that (link 1, link 2) and obviously most transfers in client workloads never touch the TLC portion until completion as writing >3GB of data at once is quite rare.

Interesting. Does the Samsung SM951 128GB M.2 PCI-E AHCI SSD have the same issue? Because I copy ISO files (after downloading) to my NAS, and those ISO files are sometimes greater than 3GB.

From the Anandtech ADATA SP550 review:
850 EVO 120GB scores 166.93 MB/s, So factoring in parallelism being halved on the future Samsung 512 Gb TLC V-NAND a 120GB SSD based on two of these dies should come out around 83 MB/s. This assuming the same controller was used.

I thought that I had read that Samsung was officially discontinuing the 120GB size of the 850 EVO? Perhaps that change was prompted by a lithography shift, and due to lack of proper parallelism in the theoretical updated 120GB drive, they discontinued it instead?

 

[cbn]

I thought that I had read that Samsung was officially discontinuing the 120GB size of the 850 EVO? Perhaps that change was prompted by a lithography shift, and due to lack of proper parallelism in the theoretical updated 120GB drive, they discontinued it instead?

Yep, with the Introduction of the 750 EVO (which uses planar TLC NAND), the 120 GB 850 EVO is out:

http://www.thessdreview.com/daily-n...-market-and-more-storage-visions-2016-update/

However, I don't think the discontinuation of the 120GB EVO is due to lack of parallelism as the 48 layer 256 Gb TLC V-NAND have the same parallelism as the 32 layer 128Gb TLC V-NAND. This due to a doubling of planes (2 vs. 1) on the 48 layer 256 Gb.

My guess is that the planar TLC NAND dies must be cheaper for Samsung to make than 48 layer TLC V-NAND 256 Gb dies.

 

[VirtualLarry]

My guess is that the planar TLC NAND dies must be cheaper for Samsung to make than 48 layer TLC V-NAND 256 Gb dies.

I thought that Samsung had claimed, somewhere, that the cost crossover point for production of 3D V-Nand was at their 48-layer version, and that V-Nand would therefore be cheaper than Planar TLC.

Guess not? I was hoping for 3D V-NAND throughout Samsung's line-up. (Well, in MLC mode for their "Pro" devices, but otherwise the same.)

 

[cbn]

I thought that Samsung had claimed, somewhere, that the cost crossover point for production of 3D V-Nand was at their 48-layer version, and that V-Nand would therefore be cheaper than Planar TLC.

Guess not? I was hoping for 3D V-NAND throughout Samsung's line-up. (Well, in MLC mode for their "Pro" devices, but otherwise the same.)

In that second link that Hellhammer provided in post #4 the article mentions Samsung using eight dies per package:

There are three different PCB designs in the 850 EVO lineup. The 120GB and 250GB models (above) use a tiny PCB with room for two NAND packages (one on each side). Interestingly enough, both use octal-die packages, meaning that the 120GB 850 EVO only has a single 128GB (8*16GB) NAND package. Decoding the part number reveals that the packages are equipped with eight chip enablers (CEs), so a single NAND package is viable since all eight dies can be accessed simultaneously.

The use of octal-die packages is actually true for all capacities. It's an interesting choice nevertheless, but I suspect Samsung's packaging technology is advanced and mature enough that it's more cost efficient to use high die count packages and small PCBs instead of larger PCBs with more and less dense NAND packages.

And as the Anantech 750 EVO News Post points out a 120 GB 850 EVO would only have four chips per package (if using the 48 layer 256 Gb dies):

The previously announced and now imminent migration to Samsung's 48-layer V-NAND won't apply to the 120GB 850 EVO, as the 256Gb per die capacity would mean building a drive with only four flash chips. That is undesirable from both a performance standpoint and from a packaging standpoint—Samsung will otherwise have no reason to stack fewer than 8 dies per package.

So maybe 48 layer TLC V-NAND is a bit cheaper than 16nm planar TLC NAND, but the special "one off" NAND packaging required for a 120 GB 850 EVO was just too much headache for them.

 

[cbn]

Regarding the 120GB 850 EVO being discontinued, Samsung does have a 128GB TLC V-NAND 2.5" drive in the works (the OEM PM871a):

http://www.anandtech.com/show/10181/samsung-readies-new-maia-ssd-controller

This drive will use the new Maia controller rather than the MGX found in the 850 EVO.