The polysilicon annulus that forms the NAND string channel contacts the selective epitaxially grown (SEG) contact that protrudes upwards from the substrate. A metal gate lying over top the substrate serves as a select gate for the adjacent sourceline. The polysilicon channel layer is surrounded by a thin tunnel dielectric that is likely formed by atomic layer deposition (ALD). A charge trap layer, typically silicon nitride, is in contact with this tunnel dielectric and covered by an oxide barrier layer. These were also likely formed by ALD. A barrier layer oxide and metal gate (word lines) surround the charge trap layer.
Here is a SEM cross-section taken through the array portion of the memory. We see 55 gate layers in the stack: 48 NAND cell layers, 4 dummy gates, 2 SSL and 1 GSL. Two V-NAND strings, each having a polysilicon annulus surrounded by a charge trap layer and metal gates, are visible between the tall tungsten filled sourceline contacts.
Here is the NAND string in plan view, confirming the annular shape to the layers. Charge trap layers are a fairly new design for Flash memories, as they typically use polysilicon floating gates to store charge. We have seen silicon nitride trap layers being used by Spansion in their MirrirBit NOR Flash memory but Samsung might be the first to use the charge trap layer in NAND Flash.
Microsoft has included 128 GB V-NAND based SSDs in their Surface Book and Surface Pro 4 laptops. We thought had found them in the universal flash storage (UFS) NAND Flash memory used in Samsung’s Galaxy S7 (as above). This 32 GB UFS 2.0 memory is claimed to have a lower power consumption and smaller form factor than its predecessor: the eMMC type memory module.
We were hopeful that the UFS NAND Flash memory used in Samsung’s Galaxy S7 might contain some 32L V-NAND, but this was not to be, as we found their 16nm planar NAND Flash instead. But with V-NAND fabs running in both South Korea and Xi’an, China, we think Samsung has deeply committed to V-NAND and the end of planar NAND scaling is at hand.
With a higher magnification image of the top portion of the V-NAND string, you can see that the top surface of the polysilicon annulus (NAND string channel) is contacted by a tungsten bitline strap that connects to the overlying bitlines. The metal word lines, oxide barrier and charge trap layers surround the polysilicon channel layer.
By comparison, Samsung’s 16nm node planar NAND flash measures about 740Mb/mm2. So while the V-NAND is fabricated at a larger process node (~21nm vs. 16nm), its memory density is almost 3.5 times better than the 16nm planar NAND flash.
One of the 256 Gb dies is shown in Figure 4 and comprises two 5.9 mm x 5.9 mm large banks of NAND Flash memory. We can calculate a gross measure of the memory density by dividing the entire die area into the memory size to get about 2,600Mb/mm2.
Samsung announced its 256 Gb 3bit multi-level cell K9AFGY8S0M 3D V-NAND in August 2015, stating that it would be used in a variety of solid state drives (SSDs), and that it would be on the market in early 2016. True to their word, we managed to find them in their 2TB capacity, mSATA, T3 portable SSD. Source: TechInsights.
Here is a package cross section showing the 16 dies stacked one on top of the other and connected using conventional wire bonding technology. The dies are an outstandingly 40µm thin, perhaps the thinnest dies that we have seen in a package. By comparison, the dies in Samsung’s 32L V-NAND that we examined in 2014 were about 110µm thick and stacked 4 dies high in their package. Other thin memory dies that we have seen include Hynix’s HBM1 memories used in AMD’s R9 Fury X graphics cards that are about 50µm thick and some 55µm thick DRAM dies in Samsung’s DDR4 with 4 stacked dies with TSV interconnections. The 40µm might be close to the thinnest that can be achieved with 300mm diameter wafers without a carrier wafer.
Tearing open the drive, we find a double-sided circuit board containing four 0.5TB capacity K9DUB8S7M packages. Each of these packages contain 16 of the 48L V-NAND dies that we were looking for. Source: TechInsights.