Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Samsung launches the 850 Pro, the first SSD with "V-NAND" cell stacking
It's long been known that "simple" NAND die shrinks (scare quotes because I understand that die shrinks are hard) can only go on for so long. Each subsequent die shrink reduces the number of writes, and quantum mechanics will eventually mean that you just can't go any further. You can't split an electron in half, and if one electron out of place means you lose data, the drive will be useless.
So what's the solution? One approach is to stack NAND cells on top of each other, sometimes called "3D NAND". I use Samsung's name for it in the title of this thread. The idea is that instead of using a "simple" 16 nm process node like Crucial recently launched the MX100 on, Samsung used a 40 nm process node--but stacked cells 32 high. This actually gives you more data per mm^2, and with the write endurance that you'd expect of a 40 nm process node.
The downside is that stacking NAND cells 32 levels high is tricky, and more expensive than a single layer. Thus, Samsung is pricing the 850 Pro as costing more than a Crucial MX100 with double the capacity. So it's not a product that I'd recommend buying just yet.
But it's still an important tech milestone, I think. Stacking NAND cells has gone from theoretical to experimental to now available at retail. If the various proposed replacements for NAND flash don't catch on soon, such 3D NAND is likely to be the future of SSDs. While Samsung charges more for the new 850 Pro than SSDs on ~20 nm nodes do, they do charge a lot less for a given capacity than older SSDs build on ~40 nm process nodes did. There's no obvious reason why they can't eventually shrink that 40 nm down to similar or slightly larger scales than single layer NAND process nodes.
Incidentally, this isn't going to be the future of CPUs or GPUs, or at least not with anything remotely similar to current techniques. You can stack NAND cells because they barely put out any heat. CPUs and GPUs put out a ton of heat. Try to stack CPU transistors 32 high like this and it would fry in a hurry. And that's ignoring that I'm not sure what stacking transistors would even mean, as there need to be metal layers in addition to silicon; transistors don't do you any good if they're not connected to anything.
Comments
I was going to ask "Isn't Haswell already on a 3D process?" It is, but it's not the same thing at all. The Haswell process and FinFETs are 3d in the sense that they actually design the transistor in 3D space, but Samsung is actually stacking (more or less) a 2D print to create a 3D grid type layout.
I agree, this is big news wrapped up in a low-key headline. Stacking will be very big, but mostly just relegated to flash and maybe cache RAM inside a CPU/GPU die (maybe in a low-density application on GPU/CPU, but probably not nearly as high, those areas don't tend to be space hogs anyway, that is mostly the cache that eats up die space).
I'm actually surprised we are seeing a commercial application this early.
I had considered addressing that, but deleted it before the final post. Intel calling their tri-gate stuff "3D" is just stupid marketing hype. Transistors always exist in three dimensions, but they're traditionally laid out on chips one deep, but thousands or tens of thousands across. What Samsung is doing with their "3D NAND" is having chips that lay out cells tens of thousands across in each of two dimensions, and then stacked 32 high in the third dimension.
For CPUs without a GPU attached, die space barely matters anymore, as the CPU cores are so small as it is. If the CPU cores you have planned use up 10 mm^2 of die space and the chip needs to be at least 100 mm^2 to have enough space for pads to connect to the motherboard, then making the cores smaller doesn't help directly. Die shrinks do tend to help with power efficiency.
For GPUs, saving on die space may well always be useful. If you have twice as many shaders clocked half as fast, you can get the same performance, but with considerably lower total power consumption. Having a large cache on the GPU die will also greatly ease up on memory bandwidth once you don't have to constantly access memory for the depth buffer or frame buffers.
Samsung better hurry. HP's memristor technology may change storage if it does everything it says it will do.
http://www.engadget.com/2014/06/11/hp-the-machine/