Review Index:
Feedback

Samsung 850 Pro Full Capacity Roundup - Capacity Independent Speed

Subject: Storage
Manufacturer: Samsung

Introduction

Given that we are anticipating a launch of the Samsung 850 EVO very shortly, it is a good time to back fill on the complete performance picture of the 850 Pro series. We have done several full capacity roundups of various SSD models over the past months, and the common theme with all of them is that as the die count is reduced in lower capacity models, so is the parallelism that can be achieved. This effect varies based on what type of flash memory die is used, but the end result is mostly an apparent reduction in write performance. Fueling this issue is the increase in flash memory die capacity over time.

View Full Size

There are two different ways to counteract the effects of write speed reductions caused by larger capacity / fewer dies:

  • Reduce die capacity.
  • Increase write performance per die.

Recently there has been a trend towards *lower* capacity dies. Micron makes their 16nm flash in both 128Gbit and 64Gbit. Shifting back towards the 64Gbit dies in lower capacity SSD models helps them keep the die count up, increasing overall parallelism, and therefore keeping write speeds and random IO performance relatively high.

Read on for the results of our full capacity roundup!

To tackle this sort of issue, Samsung took a different and significantly more future looking approach which started several years ago. Given that constantly shrinking planar flash memory introduces several issues and complications, Samsung opted to push the process vertically. Their first significant entry was a 24 layer flash die with a capacity of 128Gbit, but that was only used in a few of their datacenter models.

View Full Size

The more recent consumer addition shifted from 128Gbit to 86Gbit dies. 86 Gbit is an almost neat 2/3 of 128Gbit, so power of two capacities are still feasable so long as the die counts per package are staggered per model (more on that later).

View Full Size

Representation of Samsung's 3D VNAND in 128Gbit and 86 Gbit variants.
Image republished with permission from Schiltron Corporation.

In the above image, the red square represents the space between cells in Samsung's 3D VNAND. The other squares represent 20nm (yellow) and 16nm (blue) planar (2D) NAND. It's fairly obvious that cells must be packed together much more closely on 2D NAND. You would think that it shouldn't matter how the cells are packed so long as those cells can store the desired bits, but the packing method is the exact thing that dictates how quickly you can store those bits. Here's why:

View Full Size

When writing to NAND, adjacent cells actually interfere with each other. Electrons destined for one particular cell ight bleed into an adjacent cell. The workaround is to adjust and verify cell voltages in incremental steps.

View Full Size

Due to the structure and spacing of Samsung's 3D VNAND, this intereference is significantly lessened. This allows for fewer iterations of the Program/Verify cycle, which ultimately results in a faster write speed. While we have no specific figures on this, Samsung states the improvement is ~2x over current 2D NAND methods. This improvement, combined with the fact that they can use 50% more 86Gbit dies in an equivalent capacity SSD using 128Gbit 2D NAND, adds up to a theoretical 3x speed improvement when moving from 2D to 3D. While this does not mean your SATA 6Gbit SSD will go faster than 6Gbit, it does mean that lower capacity models will be less likely to experience slow downs. It also means that a VNAND equipped PCIe SSD would, apart from being awesome, saturate its faster link with fewer channels and dies as well.

Here's a repeat of the cool 3D VNAND video from our initial review of the 850 Pro 512GB.


November 3, 2014 | 06:46 PM - Posted by Anonymous (not verified)

These new charts are not the easiest to read.
Slightly larger text and solid colored bars would help a lot.

November 4, 2014 | 09:04 AM - Posted by willmore (not verified)

I second this. I find them very hard to read. The bar charts are okay because they are all in order, but when they wonder around like the line graphs, they're practically illegible. Add in the mislabeled axies and I'm not sure what I'm supposed to be reading.

November 4, 2014 | 08:01 PM - Posted by pdjblum

Agreed.

November 5, 2014 | 01:12 PM - Posted by kir (not verified)

Not the easiest graphs to read!
Background color hinder very much.

November 3, 2014 | 07:28 PM - Posted by Melvar

It looks like you mislabeled the X axis on the PCPer File Copy Test charts. I'm guessing it's supposed to be "seconds" rather than "MB/s".

November 3, 2014 | 08:29 PM - Posted by Anonymous (not verified)

Agreed. Also, i'm pretty sure that for the "IO Meter Average Transaction Time" lower - not higher - should be better.

November 4, 2014 | 12:16 PM - Posted by Allyn Malventano

Thanks for the catch, I've re-uploaded corrected versions, and expanded the QD out to 8 on these to more easily see where the drives are headed. Expanding it further would make the lower part of the scale harder to read.

November 4, 2014 | 08:56 AM - Posted by willmore (not verified)

Yeah, I almost went out and bought an ADATA SP610 in 128GB. :) That thing beat everyone!

November 4, 2014 | 12:05 PM - Posted by Allyn Malventano

Got it. Fixed the axis label on file creation and copy charts. Thanks for reporting this.

November 3, 2014 | 08:12 PM - Posted by Anonymous (not verified)

Look at how much extra room is available in that 2.5in casing for the 512 and 1tb models! Going with full 16 layer packages and filling up the remaining space with them could yield a 2.5 inch SSD in excess of 3TB. O_O

This tech needs to hurry up and get fully commoditized so Samsung has an incentive to make drives that size and we can actually buy them at reasonable prices. Doing so would put a few more well earned nails in the HDD makers coffins.

November 3, 2014 | 08:55 PM - Posted by Grand Nagus (not verified)

hey Allyn when are you getting the intel 750 pci-e ssd? supposed to be out in Q4 and well... its Q4 now!

November 4, 2014 | 08:44 PM - Posted by Anonymous (not verified)

Possible error with the YAPT chart. Shows queue depth on the bottom when it should be block size(or whatever the correct term would be). Regardless, interesting new benchmark along with a great article.

November 5, 2014 | 05:12 PM - Posted by Allyn Malventano

Fixed, and thanks!

November 5, 2014 | 02:31 PM - Posted by avrion

I'm wondering, can these drives perform any faster if connected to 12gb/s SAS ports?

November 5, 2014 | 05:13 PM - Posted by Allyn Malventano

The controller and connector do not support SAS 12Gbit/sec. Even if connected to an appropriate HBA, the link would fall back to 6Gbit SATA.

November 12, 2014 | 12:39 AM - Posted by DrWattsOn

As usual I'm late to posting comments, but in case anyone notices my post: I have not found any source at Samsung or here where it addresses some sort of "hold-up" protection on the power line in case of a PSU failure. Anyone know?

Post new comment

The content of this field is kept private and will not be shown publicly.
  • Lines and paragraphs break automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <code> <ul> <ol> <li> <dl> <dt> <dd> <blockquote><p><br>
  • Web page addresses and e-mail addresses turn into links automatically.

More information about formatting options

By submitting this form, you accept the Mollom privacy policy.