Micron recently added to the alphabet soup of the NAND flash storage market with the first QLC flash product to come to market. And Intel plans to launch its first QLC flash product in the second half of 2018.
QLC – or quad-level cell – is the latest in a line of flash generations that began with SLC (single level cell, or one bit per cell) and progressed through MLC (two bits per cell) and TLC (three bits per cell).
But what is QLC and what will it mean for the flash storage market?
As the name suggests, QLC allows for four 0 or 1 states per flash cell. That provides for 16 different combinations per cell, namely 0000, 0001, 0010, 0011, 0100, 0101, 0110, 0111, 1000, 1001, 1010, 1011, 1100, 1101, 1110, and 1111.
That compares with just two – 0 or 1 – in single-level cell flash and four or eight in MLC and TLC, so clearly QLC boosts the density of storage by 2x – and that’s not to be sniffed at.
But there is a downside. QLC can pack more data into a smaller area than existing flash generations, but is also vastly more sensitive to wear upon writes.
And in fact, with QLC, we are looking at media that really cannot be written and rewritten to that many times.
That’s because flash storage is based on reading the flow of voltages between different layers of material, with that actual mechanism of voltage flow explicable at the level of quantum physics.
To generalise, QLC has to do its switching and measuring in very tight spaces which are affected by wear, making reading voltages much less certain. So, writing in any flash – and even more so in QLC – literally eats away at the ability to write and rewrite data.
Micron would not give performance figures for the 5210 QLC, but provided ratios compared to its 5200 TLC Eco drives.
Read more on flash storage
- New media provide a range of options to speed workloads, from “old-fashioned” flash to storage class – and persistent memory. We help you to exploit the storage performance hierarchy.
- Flash storage is mainstream, but what do storage professionals need to know about Raid when used with SSDs? We look at Raid fundamentals, and supplier implementations.
So, for example, while random reads will come in at 0.8x to 1x of 95,000 IOPS, sequential write performance will be 0.6x to 0.8x of 520MBps.
For random writes, performance is a sluggardly 0.25x of 22,000 IOPS. That’s about 6,500 IOPS and still way more than what you would expect of a spinning disk HDD, which could clock maybe 200 IOPS per drive.
But IOPS is not the QLC drive’s Achilles’ heel. Its endurance is the main limiting factor.
The 1.92TB Micron TLC 5200 Eco has stated endurance of 3.5PB in terms of the expected lifetime volume of data writes it can take; the 5210 QLC can achieve only 0.05x or 0.1x of that.
So, you can expect to write between 175TB and 350TB to that 1.92TB drive in its lifetime. That would roughly equate to filling and refilling it with data about 180 times.
With its limitations, QLC is destined to have a quite limited set of use cases.
These will be something like those outlined by Micron, namely large-scale, perhaps webscale, operations where there is a need for rapid reads that will vastly outnumber writes, so probably data that will rarely, if ever, be rewritten.
There is a market there for those that are currently short stroking HDDs to do the job.
There question now is, with the limits of physical scale affecting operations so markedly, can any more be wrung from NAND flash, or is QLC the last generation?