Samsung Introduces TLC SSD

Samsung on Monday introduced a new “840” SSD series which reviewers have found is based on TLC flash.

Oddly enough the press release for this product seems only to have been distributed in Korea to reviewers who attended a special introduction of the device.  The SSD Guy has not been given the specifications presented at the event, and had to ask Samsung for a copy of the press release.

The press release focuses on the product’s 100,000 read IOPS, that it comes in two versions, the “Pro” model for the enterprise and another model for client applications, and the fact that the controller uses a new design based on three ARM cores. A read IOPS figure of 100,000 is very high performance for a SATA drive!  One has to wonder if the client market will be able to distinguish between this level of performance and drives with fewer than 10,000 IOPS.

Other specifications had to be found through web searches of news items and blog posts by those who actually attended the event.

What I have been able to glean from reports of the reviewers who were flown to Korea for the launch is that the product has very fast specifications.  The client drive is also the first SSD to use TLC flash.

That’s a big deal.  TLC is hard to manage and has never before been used in an SSD.  Samsung didn’t find this distinction worth including in the press release, perhaps because so few industry observers and participants believe that TLC can be used to build a reliable SSD.

Let’s explore that a little bit.

The graphic in this post has been put together to introduce readers to the difficulty of accurately reading data out of TLC flash.  Here we have a microscope laid on top of eight equal size boxes, each representing a voltage level in the flash – TLC flash stores three bits per cell by using eight voltage levels.  (Purists argue that this implies that it shouldn’t be called TLC for Three-Level Cell or Triple-Level Cell, but The SSD Guy won’t address that debate in this post.)

Sitting at your desk it is a little slow to determine which box some part of the microscope, like the focus knob, is in.  It would be quicker and the answer would be more certain if the background had half as many boxes, which would represent the four voltage levels of MLC, and your decision would be even faster (and the results even more certain) using only two boxes, one upper and one lower, like SLC flash.

Digressing a little, let’s look at what process shrinks do to this: Move across your office and try the same exercise.  Which box is the mirror in?  Now move twice as far away and try to determine which box the center of the stage is over.  Each time a new process geometry is introduced it’s equivalent to doubling your distance from the graphic.

Clearly TLC presents new challenges to flash users.  Why, then, would anyone want to design TLC into a system?  In a nutshell: it’s cheaper than MLC.  The die size of a TLC chip is 20-25% smaller than its MLC counterpart, and a smaller die size means a cheaper part.  Furthermore, consumer applications are moving to TLC flash, and that means that the economies of scale will eventually make TLC significantly cheaper than MLC.  This has occurred with SLC flash which now sells for as much as seven times the price of its MLC counterpart, even though an SLC chip is only about twice as large as the MLC chip.  This is driven by the fact that MLC accounts for over 95% of the NAND gigabytes sold today.

In this vein, every application that can migrate to TLC flash will eventually make that move.  Although many have argued that TLC will never find its way into SSDs, Samsung has proven that this is not the case.  The SSD Guy applauds Samsung on taking this bold step to lead the way into SLC SSDs.

Only yesterday I read a comment from an SSD maker arguing why TLC would never be used in an SSD.  This made me reflect on the early days of SSDs (2006-7) when MLC was deemed “unusable” in any SSD.  Once MLC-based client SSDs were introduced (2008-9) pundits argued that MLC was “unusable” in enterprise SSDs. Today MLC is common in enterprise SSDs.

The motivation to use TLC is not all that compelling today, but that could change.  Today TLC and MLC prices are close to each other, while SLC prices are about 7 times those of MLC (although common sense would lead you to expect 2x – but that’s market dynamics for you!)

Should consumer applications migrate en masse to TLC it would not be even slightly surprising for MLC to rise to 7 times the cost of TLC.

If that happens, SSD makers may find that it makes a lot of sense to replace a certain capacity of MLC with three times as much TLC at a lower total cost: they would stll save money by employing significant overprovisioning to accelerate performance and manage wear (which is discussed in some depth in The SSD Guy’s Friday series on How Controllers Maximize SSD life.)