Samsung Crams Entire 512GB NVMe SSD Into Single BGA Chip Package

Subject: Storage | May 31, 2016 - 03:38 PM |
Tagged: TurboWrite, Samsung, PM971-NVMe, BGA, 512GB, 48-layer, 32GB, 256Gbit

Have you ever checked out one of those laptops with the soldered-on eMMC SSD, where the manufacturer was basically checking the 'SSD' box for forgetting the 'Performance' box entirely? What if I told you that it was possible to fit an entire PCIe NVMe SSD with performance comparable to a 950 Pro into a package similar to those eMMC parts?

View Full Size

We have seen Samsung SSD internal PCBs shrinking over the past few revisions of their EVO line, but this is certainly taking things to the next level.

View Full Size

This new PM971-NVMe SSD is stacking 16 32GB 48-layer V-NAND dies lifted from their V2 850 EVO, stacking those *on top of* a 4GB of LPDDR4 20nm RAM die, which sits on top of Samsung's own ARM-based SSD controller. That's 768 layers of NAND, 4GB of NAND, and an NVMe controller all in a single BGA package! The complete package is rated at 1.5GB/s reads and 900MB/sec writes (SLC speed), and since everything is together and soldered straight to the mainboard of the system, thermal throttling should not be an issue since the system board would double as the heat sink. 512GB will not be the only capacity for this part - it will also be available to OEMs in 256GB and 128GB capacities (likely with lower performance specs as there will be reduced flash die counts in those parts, meaning reduced parallelism.

This is a big deal in a few different ways. Not only will it enable much higher performance in smaller mobile platforms, it is also the first manifestation of TurboWrite TLC/SLC caching with an NVMe linked controller (the 950 Pro was straight MLC). Being able to write straight to SLC does wonderful things for the 850 EVO's Latency Weighted Percentile results (bottom of page) even over a SATA bus, so I can only imagine what it would do over a PCIe link. The only potential downside would be that a laptop with PM971-NVMe installed would mean that it is not user replaceable, but then again, why would you want to?

Source: Samsung

Video News

May 31, 2016 | 04:28 PM - Posted by Tomlai (not verified)

When people thought SSD tech is going nowhere...

May 31, 2016 | 06:42 PM - Posted by Anonymous (not verified)

Who really needs a 512 Gio fast unreliable storage?

For the operating system 16 Gio should be enough.

June 6, 2016 | 01:32 PM - Posted by Allyn Malventano

What decade are you from?

May 31, 2016 | 05:40 PM - Posted by Anonymous (not verified)

I just got 512 boners.

May 31, 2016 | 06:06 PM - Posted by DieH@rd (not verified)

Nextgen consoles NEED to start using this.

May 31, 2016 | 06:12 PM - Posted by Anonymous (not verified)

One high density chip but no metallic package to ease the cooling of transistors... what a failure!

May 31, 2016 | 06:22 PM - Posted by Anonymous (not verified)

They could put the controller on top to allow for placing a heat sink, although the power consumption of most of these controllers is only a few watts at most. That doesn't need much cooling.

May 31, 2016 | 06:56 PM - Posted by Anonymous (not verified)

They could put metallic slices between transistor layers connected at tips to make a TO-220 like package.

Stacking layers without metallic package prevents the natural convection from working well and makes the storage less reliable, furthermore for high density chips (20 nm).

May 31, 2016 | 08:30 PM - Posted by Anonymous (not verified)

The actual flash die themselves do not put out that much heat. It is mostly the controller. I just looked up some power consumption numbers for a Samsung 850 evo 1 TB under sequential write is only 2.61 watts. The highest was an old 840 evo 1 TB at around 4 watts. I don't know how high the 950 is, but it is probably quite low also. Even if the controller is on the bottom of the stack, a 4x pci-e link takes up a relatively small number of pins. They could transfer some heat out through the socket.

May 31, 2016 | 08:53 PM - Posted by Anonymous (not verified)

Do you mean Flash cells do not use switching transistors at all?

Only the controller needs transistors to work?

What about the FGMOS array?

I doubt SSD automagically run without current leaks in Flash cells.

Actually you don't need a nuclear plant to make things warm.

Even 1 watt could be enough to burn your retina.

June 3, 2016 | 06:47 PM - Posted by Anonymous (not verified)

So you think a memory chip is going to consume hundreds of watts or something? CPUs and GPUs can produce hundreds of watts in a smaller die size; memory chips do not. The flash die will produce some heat, but there are few actually switching transistors on the die at any one time; you are not reading or writing an entire chip at once. With a CPU or GPU, almost the entire chip is active simultaneously. The transistors on the flash die also operate very slowly in comparison to logic die like CPUs or GPUs.

The controller is a logic die, but it is a very low power logic die. There isn't that much processing power required to control NAND die. Most of the power will be consumed by the controller, and for Samsung devices, that is only about 2 to 4 watts for the entire drive (controller plus flash) while it is actively writing. It is in the milliwatt range when idle. That amount of power can be handled without even needing a heatsink in most cases; you would have to have a very small die or very harsh operating conditions before more cooling would be required. This device is probably even more heavily optimized for low power than Samsung's high performance desktop products. The heat from stacking die is a consideration, but it is not that much of a problem for most memory devices. It is when you want to incorporate a high power GPU or CPU that stacking runs into serious heat issues.

May 31, 2016 | 09:50 PM - Posted by Anonymous (not verified)

Samsung is the first company in the world to mass produce TLC flash, the first to mass produce vertical flash (now the second generation with 48 layers), the top company in SiP (system in package, putting many dies in a package), they have the best ssd controller, they have the most reliable ssds, they are one of the top RAM companies... So for everything in that chip they are the best or one of the best companies in the world. Sure they know how to package it. ¿Or maybe your comment falls under some kind of Poe's law?

June 1, 2016 | 03:33 AM - Posted by Anonymous (not verified)

You know even the best make mistakes... ;-)

Should I recall you where came from the subprime mortgage crisis?

June 2, 2016 | 08:24 PM - Posted by Anonymous (not verified)

that was not a mistake.

June 3, 2016 | 03:32 AM - Posted by Anonymous (not verified)

It is a mistake like multicore processors if you are unable to correct the drawbacks of the technology...

June 3, 2016 | 07:08 PM - Posted by Anonymous (not verified)

Even Samsung's older 850 evo drives are less than 3 watts for the entire drive while writing. The specs say 3.7 to 4.4 watts for the 850 evo, but Anandtech measured less than 3 under sequential and random access for the 1 TB model and less than 2 watts for the 500 GB model. This new device is probably even lower power. If it uses TSV, it could be ridiculously low power. That much power in that die area shouldn't even need a heatsink.

June 3, 2016 | 07:19 PM - Posted by Anonymous (not verified)

The 950 pro goes up to around 5 watts, but even that shouldn't need a heatsink. I would expect the new device to be more power optimized, especially if it uses TSV. I would not buy a system with soldered on, non-replaceable SSD though.

June 6, 2016 | 01:33 PM - Posted by Allyn Malventano

With the chip soldered directly to a system motherboard, heat will dissipate through the BGA itself and have no issue dissipating through the board.

May 31, 2016 | 08:41 PM - Posted by Anonymous (not verified)

Is this a TSV based device then? For small form factors, it would be nice to get stacked DRAM and flash or X-point on an m.2 style device. You could make a relatively large memory module with A small jumper of stacks of DRAM on an m.2 style device. There isn't much reason to have even something as large as a SO-DIMM. Unfortunately this will probably lead to systems with SSD and DRAM soldered onto the board rather than a new expandable form factor. Given the limited lifetime of SSDs, soldering one on a board seems like a bad idea.

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

This question is for testing whether you are a human visitor and to prevent automated spam submissions.