Calling QLC junk gets half of it right. Each cell’s fourth bit costs endurance, program time, and room between voltage states. On my bench that cost shows up in one place, which is writing hundreds of GB in one go after the cache has filled. In my first set of drives, the QLC drives I could identify sat inside the range of the TLC drives on the everyday tests. A second set from my work bench shows the same after-cache gap. It also shows a lower queue depth 1 random read for QLC, which the first set didn’t. This page is one of the Flash Myths, Tested series.
Where the myth comes from
QLC stores four bits in each cell, which means 16 voltage states where TLC has 8. A bad drive is the clearest case I have for the myth. A Tom’s Hardware review of a budget NVMe line was updated after its maker swapped TLC flash for QLC under the same model name. After its cache filled, the QLC version averaged 40MB/s while the TLC version had averaged about 450MB/s. It had the larger cache, about 135GB against 24GB, and the lower burst speed, 1.16GB/s against about 1.85GB/s. Tom’s Hardware puts part of the cause down to the smaller number of flash packages, because fewer packages leave less to interleave across. Figure 1 sets the two versions side by side.
Where the myth comes from
The QLC version was slower, even with the larger cache
One budget NVMe model before and after its maker swapped TLC flash for QLC
That’s the worst case in my sources, and it’s a real one. It’s also a single drive, and the package count was part of the cause as well as the flash type.
What QLC gives up on paper
Micron’s technical note on moving from TLC to QLC is the clearest statement I have saved, and it is written for engineers deciding whether to switch. Against a similar TLC device, QLC has about 75% of the specified endurance. Its longest array read is about 2x as long and its longest array program about 3.5x as long. It stores 33% more bits per cell by doubling the states, from 8 to 16, and Micron says the closer states raise the raw bit error rate, how often error correction triggers, or both, but gives no number for it.
What QLC gives up on paper
QLC gains bits and gives up endurance and time against TLC
QLC as a share of a similar TLC part, with the TLC level marked
QLC has real gains too. In Micron’s example, 2TB of NAND takes 16 QLC dies of 1Tb or 32 TLC dies of 512Gb, half as many. In the Shibata paper, a 1.33Tb QLC die and a 512Gb TLC die built on the same 96 layers reach 8.5Gb/mm² and 5.95Gb/mm², which is 43% more bits per area. Figure 2 puts both sides on one axis.
How big the endurance gap is depends on who you ask. A 2019 paper (Liang and colleagues) summarizes rated program/erase cycles in a table, giving 3,000 to 5,000 P/E cycles for TLC and 500 to 1,000 P/E cycles for QLC, so the matching ends of the ranges are five to six times apart. That table summarizes specifications and measures nothing, and it doesn’t agree with Micron’s like-for-like 75%. Flash Foundations Part 2 labels its cycle counts illustrative because the sources disagree. I use Micron’s figure for what a current part gives up and the paper’s table only for the direction.
What my bench shows
I have two sets of drives, and I keep them apart because they were tested on different platforms, in different runs and with different kinds of sample. Set A is the test archive, covered in this section and in the pair that follows. Set B is a second set of drives from my work bench, and it comes after that pair.
The limit is that only 29 of the 86 retail drives in my archive have a confirmed cell type, 25 TLC and 4 QLC, from flash-ID dumps or from what the review stated. I left out the other 57 rather than guess. So this is a few named drives rather than a population, and I give ranges and counts below instead of averages for QLC.
Every drive with a run took the sustained-write test I describe in the speed on the box, which writes sequential 1MiB blocks at a queue depth of 32 and starts from an empty drive. Nine of the 25 TLC drives have no such run, so the TLC side below uses the other 16. Table 1 has the four QLC drives.
| Drive | Interface | Burst write | Written at burst speed | Average after the drop |
|---|---|---|---|---|
| Intel SSD 670P 2TB | PCIe 3.0 | 2,886MB/s | 269.9GB | 407MB/s |
| Kingston NV2 2TB | PCIe 4.0 | 2,971MB/s | 564.1GB | 250MB/s |
| PNY LX3030 2TB | PCIe 3.0 | 3,044MB/s | no drop in 15 minutes | none |
| Team Group T-Force Cardea Z44Q 4TB | PCIe 4.0 | 4,010MB/s | 1,045.5GB | 311MB/s |
Three of the four fell off a cache cliff and averaged 250 to 407MB/s afterward. Only the PNY LX3030 held about 3,000MB/s for its whole 15-minute run, and it wrote 2,790GB, which is more than its own capacity. It has no after-cache figure, and I can’t tell from that run what it would do on a longer one.
What my bench shows
Three QLC drives sit at the slow end after the cache
Average write speed after the cache: 16 TLC drives and 3 QLC drives
Figure 3 sets those three beside the 16 TLC drives that had a run and a cache drop. Those TLC drives span 197 to 3,299MB/s. Of the 16, 15 averaged more than the best QLC drive, 13 averaged more than 900MB/s, and 1 averaged less than the slowest QLC drive. The QLC drives had large caches and wrote 269.9 to 1,045.5GB at burst speed before the drop, against a TLC median of 213.1GB and a TLC range of 24.2 to 730.5GB. A larger cache puts the drop later but doesn’t change the speed after it.
On the everyday tests in my suite, the QLC drives weren’t separated from the TLC ones. PCMark 10 Storage scored the four QLC drives from 1,582 to 2,761 and the 25 TLC drives from 1,354 to 3,631. Of the TLC drives, 15 scored inside the QLC span, 9 scored above every QLC drive and 1 scored lower than the worst. At queue depth 1, the 4KiB random read ran 65.0 to 86.0MB/s on the QLC drives and 55.8 to 97.3MB/s on the TLC drives. PCMark replays traces of real use and reports a score instead of a time, so it only indicates how an application would run.
The one pair on the same controller
Cell type isn’t the only thing that differs between drives, so the cleanest comparison I have is a pair from one product line. It’s a 1TB TLC drive and a 2TB QLC drive, both on the SM2267XT controller, both without DRAM and using a 64MB host memory buffer.
| Kingston NV2 1TB | Kingston NV2 2TB | |
|---|---|---|
| Cell type | TLC | QLC |
| NAND | Toshiba 112-layer BiCS5 | Intel 144-layer |
| Burst write | 2,682MB/s | 2,971MB/s |
| Written at burst speed | 201.1GB | 564.1GB |
| Average after the drop | 512MB/s | 250MB/s |
| 4KiB random read, QD1 | 70.8MB/s | 70.4MB/s |
| 4KiB random write, QD1 | 277.3MB/s | 277.0MB/s |
| PCMark 10 Storage score | 2,479 | 2,528 |
| 3DMark Storage score | 2,711 | 2,718 |
The one pair on the same controller
On the same controller, QLC trails only after the cache
2TB QLC drive as a share of the 1TB TLC drive
Figure 4 divides one drive by the other. Here the QLC drive bursts 111% as fast, wrote 281% as much before dropping, and averaged 49% as fast after the drop. Its random read and both benchmark scores were within 2% of the TLC drive.
Three things differ besides cell type, and each could move the numbers. Compared with the TLC drive, the QLC drive has twice the capacity, its NAND comes from a different maker and generation, and its run was longer, 4,530 seconds against 1,542, which changes what “average after the drop” covers. So this is one pair, and one pair can’t show that QLC halves anything. It does show a shape, with the same controller, the same everyday scores and much lower sustained writes.
A second set of drives from my work bench
Set B is 39 drives from my work bench, and it’s a different kind of set. Of them, 28 are TLC and 10 are QLC, and I could not identify the cell type of one. It includes 6 TLC and 4 QLC drives that are pre-release samples, which I keep in these numbers and flag. Cell type and DRAM come from controller spec sheets, a spec table, or a folder label. Two drives, both TLC, have a flash-ID dump, and no QLC drive in set B has one. Twelve drives have an unknown DRAM state, so they are in no group below. I ran the same sustained-write test as for set A, from an empty drive, and I don’t pool the two sets in one statistic.
| Group | Drives | With a run | Median after the drop | Range after the drop | 4KiB random read, QD1 (median) |
|---|---|---|---|---|---|
| TLC with DRAM | 15 | 12 | 2,899MB/s | 876 to 3,683MB/s | 92.4MB/s (6 drives) |
| TLC, DRAM-less | 5 | 5 | 757MB/s | 108 to 1,708MB/s | 86.1MB/s (3 drives) |
| QLC with DRAM | 5 | 5 | 153MB/s | 62 to 296MB/s | 46.3MB/s (5 drives) |
| QLC, DRAM-less | 2 | 2 | 111MB/s | 95 to 127MB/s | 72.6MB/s (2 drives) |
Table 2 is the after-cache story again, and Figure 5 plots its four medians. All 10 QLC drives dropped off their cache and averaged 62 to 296MB/s afterward. Of the 24 TLC drives with a run, 23 averaged more than the best QLC drive and 1 averaged less, at 108MB/s on a 250GB drive. Medians suggest that cell type separates the groups more than DRAM does, though each group is only 2 to 12 drives and the samples are mixed. This table includes the pre-release D28, a 1TB DRAM-less QLC drive, so the DRAM-less QLC group is 2 drives at a median of 111MB/s. The DRAM-less SSDs page leaves that drive out and counts one such drive, at 127MB/s. The DRAM-less TLC range, 108 to 1,708MB/s, also overlaps the QLC ranges. TLC drives with DRAM held 2,899MB/s and the DRAM-less ones 757MB/s, while the QLC drives held 153MB/s with DRAM and 111MB/s without it. The five QLC drives with DRAM are five capacities of one product line on a PCIe 3.0 controller, which makes that group a single design.
A second set of drives from my work bench
In set B, the QLC groups sit far below the TLC groups after the cache
Set B: median average write speed after the cache, by group
Random read is where set B differs from set A. QLC drives read 4KiB at queue depth 1 more slowly, at 46.3MB/s for the PCIe 3.0 QLC drives with DRAM against 92.4MB/s for the TLC drives with DRAM. Part of that is the PCIe 3.0 link, so the cleaner check is the pair below.
The pair on one controller in set B
Set B has one same-controller pair, made of a TLC drive and two QLC drives on one controller family, all three M.2 2230 on PCIe 4.0 and all DRAM-less. The TLC drive is a mass-production sample, the 1TB QLC drive is a pre-release sample, and the 2TB QLC drive is a retail sample.
| 1TB TLC | 1TB QLC | 2TB QLC | |
|---|---|---|---|
| Sample | mass production | pre-release | retail sample |
| Burst write | 4,833MB/s | 3,973MB/s | 3,854MB/s |
| Written at burst speed | 76.2GB | 254GB | 491.8GB |
| Average after the drop | 1,545MB/s | 95MB/s | 127MB/s |
| Run length | 1,999 seconds | 4,530 seconds | 8,130 seconds |
| 4KiB random read, QD1 | 86.1MB/s | 73.3MB/s | 71.9MB/s |
| 4KiB random write, QD1 | 327.3MB/s | 448.4MB/s | 398.1MB/s |
| PCMark 10 Storage score | 2,631 | 2,874 | 2,860 |
| 3DMark Storage score | 3,286 | 3,287 | 3,168 |
Those two QLC drives averaged 6% and 8% of the TLC drive after the drop, and their burst write was 82% and 80% of it. Their random read at queue depth 1 was 85% and 84% of the TLC drive, and their PCMark 10 scores were 109% of it. Set A’s pair didn’t show the lower random read, and this one does, so I treat the random read as a difference between these samples rather than a QLC rule. The same cautions apply as for set A, because the capacities differ, the QLC runs wrote 0.68 and 0.75 times their capacity against 3.14 times for the TLC drive, and two of the three drives aren’t retail units.
What my data cannot say
Endurance is the first gap. Every drive in set A was new, with a median of 42 power-on hours when I tested it, so that set has no wear data at all, and I use no wear data from set B here either. The endurance facts on this page are ratings and a vendor’s statement. The 4TB Team Group drive is rated 800TBW with a five-year warranty, which is 200 full drive writes, or about 438GB written every day for the whole warranty. For scale, the median drive of the 86 had 16.5TB written by the end of my suite, sustained-write runs included, about 2.1% of that rating. Figure 6 draws both on one axis. That shows the rating is large next to a review. It says nothing about how fast any one drive wears.
What my data cannot say
My whole test suite used little of one drive’s rating
Terabytes written: a maker’s rating beside my own SMART readings
I’m also not comparing how quickly QLC and TLC caches recover while idle. The three QLC runs with a drop wrote 0.3 to 0.8 times their drive’s capacity, while 9 of the 16 TLC runs wrote at least their drive’s capacity, and recovery depends on that.
In set A there are three drives with an after-cache number and one pair, and cell type came from flash-ID dumps or from what the review stated. In set B the cell type of every QLC drive rests on a spec sheet, a spec table, or a folder label, and the DRAM-less QLC group is two drives. There are no application timings, only replayed traces.
What to do with this
If you move hundreds of GB at a time, the after-cache speed is the number that matters, and that applies to a video project, a backup, or a drive clone. The QLC drives I measured in both sets sat at the slow end. Look for a review that writes past the cache and reads that figure, because the box speed won’t tell you. Figure 7 lists four situations and what to do in each.
What to do with this
Judge a QLC drive by the job you give it
Four situations and what this guide says to do in each
My PCMark tests didn’t separate the QLC drives from the TLC ones for an operating system drive, a game library, or everyday work. Set B did show a lower queue depth 1 random read for QLC. Micron’s example shows the density QLC buys, half as many dies for the same 2TB. Compare the rated TBW at the same capacity, since Micron’s 75% is a statement about like devices. And if a drive changes its flash under the same model name, treat it as a different drive and check which version you are buying.
