Why We Need 3D NAND

For years, it has been known that traditional NAND (i.e. 2D NAND) is running on its last legs. Many analysts predicted that we would not see NAND scaling below 20nm because the reliability would simply be too low to make such a small lithography feasible. However, thanks to some clever engineering on both hardware and firmware sides NAND has scaled to 15nm without any significant issues but now the limit has been reached for real. To understand the limits of 2D NAND scaling, let's say hello to our old good friend Mr. N-channel MOSFET.

Unfortunately the diagram above is a bit too simplified to truly show what we need, so let's look at a real cross-section photo instead:

Let me walk you through the structure first. At the top is the control gate, which is a part of a structure known as a wordline. In a standard NAND design the control gate wraps around the floating gate and the two gates are separated by an insulating oxide-nitride-oxide layer (i.e. ONO), which is sometimes called Inter Poly Dielectric (IPD). Under the floating gate is the tunnel oxide, which is also an insulator, followed by the silicon substrate that acts as the bitline. 

The reason why the control gate is wrapped around the floating gate is to maximize capacitance between the two. As you will soon learn, the capacitance between the gates is the key factor in NAND as it means that the control gate can control the floating gate.

The purpose of bitlines and wordlines can be rather difficult to understand when looking at a cross-section, so here is what it all looks like from the top. Basically, bitlines and wordlines are just lines going in perpendicular directions and the floating gate and other materials reside between them.

When programming a cell, a high voltage of around 20V is applied to the wordline of that cell. Of course, the problem is that you cannot apply voltage to just one cell because the whole wordline will be activated so in order to select a specific cell, the bitline of that cell is held at 0V. At the same time, the neighbouring bitlines are charged to about 6V because this increases the capacitance between the bitline and floating gate, which is turn negates enough of the capacitance between the control and floating gate that the electrons cannot tunnel through the tunnel oxide. This is crucial because if all the bitlines were held at 0V, then all the cells along that wordline would be programmed with the same value. 

To erase a cell, a reverse operation is performed by keeping the wordline at 0V while issuing a ~20V voltage on the bitline, which makes the electrons flow in the opposite direction (i.e. from the floating gate back to the bitline/silicon).

The way NAND is programmed and erased is also its Achilles' Heel. Because such high voltage is needed, the insulators around the floating gate (i.e. ONO and tunnel oxide) wear out as the NAND goes through program and erase cycles. The wear out causes the insulators to lose their insulating characters, meaning that electrons may now be able to escape the floating or get trapped in tunnel oxide during a program or erase. This causes a change in the voltage state of the cell.

Remember, NAND uses the voltage states to define the bit value. If the charge in the floating gate is not what it is supposed to be, the cell may return an invalid value when read. With MLC and TLC this is even worse because the voltage states are much closer to each other, meaning that even minor changes in the voltage state may cause the voltage state to shift from its original position, which means the cell value will also change. Basically, MLC and TLC have less room for voltage state changes, which is why their endurance is lower because a cell that cannot hold its charge reliably is useless.

Now that we have covered the operation of NAND briefly, let's see what this has to do with scaling. Here is the same cross-section as above but with some dimensions attached.

That is what a cross-section of a single cell looks like. When NAND is scaled, all these dimensions get smaller, which means that individual cells are smaller as well as the distance between each cell. The cross-section above is of IMFT's 25nm NAND (hence the bitline length of 25nm), so it is not exactly current generation but unfortunately I do not have any newer photos. There is no general rule to how much the dimensions shrink because 16nm simply means that one of the lengths is 16nm while others may not shrink that much.

The scaling introduces a variety of issues but I will start with the cell size. As the cell size is shrunk, the size of the floating gate is also shrunk, which means that the floating gate is able to hold less and less electrons every time the process node gets smaller. To put this into perspective, Toshiba's and SanDisk's 15nm NAND is stores less than 20 electrons per NAND cell. With TLC, that is less than three electrons per voltage state, so there is certainly not much headroom for escaped electrons. In other words, the cell becomes more vulnerable to the IPD and tunnel oxide wear out because even the loss of one electron can be vital to the voltage state. 

The second issue is the proximity of the cells. The key factor in NAND is the capacitance between the control and floating gate but as the cells move closer to each other through scaling, the neighboring cells will also introduce capacitive coupling. In simpler terms, the neighboring cells will interfere more as the distance between the cells shrinks. The obstacle is that the interference varies depending on the charge of the neighbouring cell, so there is no easy way to exclude the intereference. This in turn makes programming harder and more time consuming because a higher voltage will be needed to achieve the sufficient capacitance between the control and floating gate to make the eletrons tunnel through the oxide. 

The graph above outlines historic rate of how cell to cell intereference increases through die shrinks. At 90nm, the interference was only around 8-9% but at 20nm it is a rather significant 40%. The interference means that 40% of the capacitive coupling comes from the other cells, making it very hard to control the gate you are trying to program or read. Fortunately as a result of some clever engineering (i.e. an airgap between the wordlines), the intererence is only about 25% at 25nm, which is much more managable than the 40% the historic rate would have given us. 

The above can be fairly tough to digest, so let's do a simple analogy that everyone should be able to understand. Imagine that you have a bunch of speakers with each playing a different song. When these speakers are relatively large and far away from each other, it is easy to properly hear the song that the speaker closest to you is playing. Now, what happens if you bring the other speakers closer to the speaker you are listening? The other speakers will interfere and it becomes harder to tell your song apart from the others. If you turn down the volume or switch to smaller speakers with lower output volume, it becomes even harder to distinquish your song from the songs that the other songs that other speakers are playing. If you repeat this enough times, there will be a point when you are hearing your song as unclearly as the other songs.

The effect is essentially the same with NAND scaling. When the cells, or speakers in the analogy, move closer to each other, the amount of interference increases, making it harder to sense the cell or listen to the speaker. At the same time the size of the cell (or speakers) is shrunk, which further complicates the focus on one cell (or speaker). 

That is NAND scaling and its issues in a nutshell. We have seen innovations such as airgaps between the wordlines to reduce cell-to-cell interference and a high K metal gate instead of a traditional ONO IPD to increase control gate to floating gate capacitance, but the limit has now been reached. However, like other semiconductors NAND must follow the Moore's Law in order to get more cost efficient. If you can no longer scale in the X and Y dimensions, what do you do? You hit the reset button and introduce the Z dimension.

Introduction, The Drives & The Test 3D NAND: How It Works


View All Comments

  • TrackSmart - Tuesday, July 1, 2014 - link

    I second this. The Anandtech SSD tests were designed so that we could tell the difference between drives that are all so fast - that there is no way to tell them apart in ordinary usage scenarios. I see the value of testing the theoretical performance of drives as manufacturers push the technological limits.

    That said, at the end of the day user-experience is what matters. I agree with emn13 that the "light workload" test is already more strenuous than anything the average user is likely to do, and looking at the chart, we see that almost every drive is within a range of ~280 to ~380 MB/s. I'm guessing that the range in performance gets even narrower for "real world" workloads.

    So keep up the innovative SSD testing, but be sure to put these theoretical performance gains into a real-world context when you get to the Conclusions section of these articles. Not everyone will benefit from these theoretical increases in performance.
  • hojnikb - Tuesday, July 1, 2014 - link

    Is Samsung planning on doing TLC based V-NAND anytime soon ?
    It would be great for a mainstream drive, since endurance would be higher (due to older node), speeds would probobly also went up (so no need for gimicks like turbowrite).
    Or is it not mature enough to scale down to TLC ?
  • artifex - Tuesday, July 1, 2014 - link

    You had me at 10 years warranty. I don't mind the slight premium if I'm not buying another one midway through the cycle. Sure, it will be obsolete well before it dies, but that term signals Samsung is really confident about their reliability. Reply
  • Gigaplex - Tuesday, July 1, 2014 - link

    Since it's twice the price of competition like the MX100, you're better off replacing mid way through the cycle. Reply
  • Arnulf - Tuesday, July 1, 2014 - link

    I must have missed this in the article - are these V.NAND cells as used in 850 Pro drives 2 or 3 bits per cell ? I got the "larger lithography improves endurance" part, I'm just wondering whether they opted for more conservative option (MLC) there as well. Reply
  • extide - Tuesday, July 1, 2014 - link

    These are MLC, or 2 bit per cell.

    It would be interesting if the non pro 850 comes out with TLC V-NAND!
  • himem.sys - Tuesday, July 1, 2014 - link

    Heh, we are waiting for tests 850pro vs 840pro, because there are no bigger differences "on paper". Reply
  • sirvival - Tuesday, July 1, 2014 - link

    one question:
    In the review the idle power consumption for e.g. the 850 128gig is 35 mw.
    I wanted to compare that to my Samsung 470 so I went to Bench and selected the drives for comparison.
    There it says that the 850 uses 0.29 Watt.
    So how comes there is a difference?
  • KAlmquist - Tuesday, July 1, 2014 - link

    Anandtech Bench has four SSD power numbers:
    SSD Slumber Power (HIPM+DIPM)
    Drive Power Consumption - Idle
    Drive Power Consumption - Sequential Write
    Drive Power Consumption - Random Write

    The confusing things are that (1) the review only listed slumber power, not idle power, and (2) Bench lists both numbers but doesn't place the slumber power next to the other power values.
  • mutantmagnet - Tuesday, July 1, 2014 - link

    I also find the lack of powerloss protection being a big negative over this hard drive. Until REFS has all the features it needs in Windows that you would get in Linux this is going to be an important feature for anyone who values data integrity. Even after that happens it still might be very important. Reply

Log in

Don't have an account? Sign up now