Is a QLC Solidigm D5-P5336 a sensible drive for a write-heavy workload?

1. 1.Start on the server that runs the workload today, against the disks it already uses: the QLC drive is not installed yet, so the existing disks are the only record of what the workload writes. On an NVMe disk, read the host write counter, wait a known number of days, and read it again.

   nvme smart-log /dev/nvme0   # record Data Units Written, wait N days, read again

   ReadingThe field prints as 'Data Units Written' or data_units_written. It counts host data in thousands of 512-byte units, so one unit is 512,000 bytes and metadata is left out. Bytes written per day = (later reading - earlier reading) x 512,000 / N. The counter holds what the host sent and nothing the drive adds internally, so it shows the floor of what a QLC drive would absorb, not the ceiling.
2. 2.Turn that daily figure into drive writes per day for the capacity you are considering: divide bytes written per day by the candidate's capacity in bytes. The same daily volume is a smaller fraction of a 122.88TB drive than of a 30.72TB one, so do this per capacity.
   ReadingThe result is a DWPD figure for your workload, but it is only comparable with a rating measured on a similar write pattern. KIOXIA's own definition ties one drive write per day to a specified workload and a specified lifetime. The fourth path lists the bases Solidigm's figures sit on.

1. 1.Volume is half the answer; write size is the other half, because Solidigm says frequent small random writes tax NAND more heavily. Sample the current disks at one-minute intervals while the workload runs.

   iostat -x -d 60   # on the server running the workload today, against its current disks

   ReadingSkip the first report, which averages since boot. w/s is write requests per second after merges, wkB/s is write throughput, and wareq-sz is the mean write size in KiB (older sysstat prints avgrq-sz instead, in sectors, with reads and writes combined). A wareq-sz at or below 4 KiB alongside a high w/s is the small-random pattern; a large wareq-sz with steady wkB/s is streaming.
2. 2.An average hides the mix: a stream of 1 MiB writes plus a flood of 4 KiB journal writes can average out to a size that describes neither. Trace individual block I/Os over a representative period.

   biosnoop   # bcc tool, e.g. /usr/share/bcc/tools/biosnoop; capture a representative period

   ReadingOne line per block I/O. The T column carries R or W and the BYTES column the size of that I/O. Count the W lines by BYTES to get the distribution. The writes that matter for QLC are those smaller than the drive's preferred write granularity, found in the next path, because they cannot be aligned to its indirection unit.

1. 1.We found no Solidigm document stating the D5-P5336's indirection unit size, so read what the drive itself reports. This needs the drive in hand: an evaluation unit, or one already in the server. Solidigm states its high-density QLC drives use indirection units larger than the typical 4KiB.

   nvme id-ns /dev/nvme0n1 -H | grep -iE 'npwg|npwa|nows|optperf'; cat /sys/block/nvme0n1/queue/minimum_io_size /sys/block/nvme0n1/queue/optimal_io_size

   Readingnpwg is zero-based in logical blocks, so preferred write granularity = logical block size x (npwg + 1). When the namespace sets OPTPERF, Linux copies that value into minimum_io_size and NOWS into optimal_io_size; without OPTPERF, use the id-ns arithmetic, not minimum_io_size. On Level1Techs, the owner of an earlier D5-P5316 measured about 377MB/s at a 4KB allocation unit and about 750MB/s at 64KB, and a reply showed the drive reports a 64KiB preferred write granularity; that is one forum report about a different drive, not a P5336 figure. Solidigm rates that earlier drive's endurance on 64KB random and sequential writes.
2. 2.Set the granularity beside the write sizes from the second path.
   ReadingSolidigm's CSAL article explains the cost: when a write is not aligned to the indirection unit, the firmware must read, modify and write the whole unit, so one host write becomes several NAND operations. That shows up as write amplification and lost performance, and it never appears in Data Units Written, which counts host bytes only.

1. 1.The product brief, dated 24 June 2023, gives an IU-aligned endurance figure calculated for a 100% random-write 16KB workload, and a footnote in the brief reads 'Endurance for 100% 16K RW for 61.44TB.' The figures themselves sit in images our reader could not extract. A separate Solidigm article states that the D5-P5336 is rated at 100% 16K random writes while its TLC D5-P5430 is rated at 100% 4K random writes.
   ReadingTwo drives rated at different block sizes cannot be compared figure for figure. A P5336 rating is a 16KB-random-write number; set it against a 4KB-rated TLC figure only with that difference printed next to both.
2. 2.Solidigm's October 2023 article on QLC in mainstream workloads states the D5-P5336 is rated up to 213 PBW on sequential writes and 65 PBW on random writes, which it converts to about 0.58 random drive writes per day. The paragraph names no capacity and no block size. The article's notes separately state that maximum endurance is based on 100% 16K random writes at 61.44TB, but no marker ties that note to these figures.
   ReadingThe gap between the sequential and random figures is the point: the same drive is rated for over three times as much data written sequentially as randomly. A forum reply on ServeTheHome makes the same argument from the other side, that a random-write rating does not describe sequential rewrites of large files.
3. 3.The current product page claims that the harshest random-write workload, run 24/7 for five years at 100%, leaves the 122TB drive with capacity to spare. A parenthetical on the page bases that on 100% 4KB random writes at 18K IOPS for the 122TB capacity, which it attributes to product specifications. We did not read a specification sheet that states the 18K IOPS figure.
   ReadingThat claim is bounded by the drive's own random-write rate at 4KB: it says the drive cannot be worn out by writing as fast as it accepts 4KB random writes, not that it suits any volume you send it. As rendered to our reader, the page's Endurance Rating, Sequential Rating and Random Rating fields were blank. Side by side: (1) the product brief, IU-aligned, 100% random writes at 16KB, with a footnote naming 61.44TB; (2) the October 2023 article, 213 PBW sequential and 65 PBW random, about 0.58 random DWPD, with no capacity or block size in the paragraph; (3) the product page, the 122TB capacity, 100% 4KB random writes at 18K IOPS, 24/7 for five years. These are three different bases and we do not combine them.
4. 4.For a second vendor's view of the same mechanism, KIOXIA defines the indirection unit as the minimum unit of the FTL mapping table, measures LC9 random writes at each model's own IU size, and prints LC9 endurance at two block sizes on its 16 KiB-IU models: 0.075 DWPD at 4 KiB and 0.3 DWPD at 16 KiB. The LC9 table rendered flattened on the page we read, so that pairing follows page order, and we do not attach it to any LC9 part number. Micron places its 6600 ION QLC drive in the capacity-focused, read-intensive endurance class.
   ReadingOn KIOXIA's own figures, the same QLC drive is rated for a quarter of the drive writes at 4 KiB that it is rated for at its 16 KiB unit.
5. 5.JEDEC's JESD219 defines standard endurance-rating workloads per SSD application class, used with JESD218, the endurance requirements and test method. We did not read the workload definitions themselves, which sit behind registration. To model your own pattern instead, Solidigm's SSD Endurance Estimator takes workload type, block size, read percentage, queue depth, duty cycle and a write amplification factor, and returns estimated life and drive writes per day.
   ReadingEnter the write sizes from the second path and the read share from iostat, not a default. The estimator says it calculates from Solidigm's own workload assumptions, and the page we read showed no D5-P5336-specific data.

1. 1.This test writes straight to the raw namespace and destroys all data on it, permanently. Run it only on an empty evaluation drive, never on one that holds data. fio refuses jobs it considers destructive against a mounted device, which is a safeguard, not a reason to point it at a live disk.
   Destructive. This command destroys what is on the device and cannot be undone. Confirm the device path twice and have a backup of anything you still need before running it.

   fio --name=wl --filename=/dev/nvme1n1 --rw=randwrite --bs=<workload write size, e.g. 4k or 16k> --ioengine=libaio --direct=1 --iodepth=32 --numjobs=4 --time_based --runtime=3600 --group_reporting   # TEST DRIVE ONLY: overwrites the device

   Reading--ioengine=libaio is there because fio otherwise uses the synchronous psync engine, on which an iodepth above 1 has no effect. Set bs to the write sizes measured in the second path (bssplit= for a mix, rw=write for a sequential stream). Run long enough to pass any burst buffer and judge the rate it settles at, not the opening rate. Read Data Units Written before and after to confirm the bytes the test wrote.
2. 2.Compare the settled rate with what the workload needs at its peak, and repeat at the drive's preferred write granularity from the third path to see how much aligned writes recover.
   ReadingIf the settled rate at your real write size covers the peak and the daily volume sits inside the rating on a matching basis, the drive fits. If only the aligned run passes, the fix is changing write sizes, not the drive.

We found no Solidigm document saying QLC is unsuitable for any workload. The documents we read state where it fits: QLC as an ideal fit for read-heavy workloads and value for workloads that are not random-write-intensive, TLC as a good fit for mixed and write-heavy ones.

Not directly. The product brief dates from June 2023, when its footnote says only 15.36TB and 30.72TB U.2 and E1.L drives were shipping, and it describes capacities up to 61.44TB. The 122TB endurance claim comes from the current product page and rests on a 4KB random-write basis stated in a parenthetical on that page.

Solidigm's Data Center Family Limited Warranty, dated 6 June 2025, runs five years from purchase in original sealed packaging, for the Original Purchaser only, and is immediately terminated when Percentage Used reaches or exceeds 100 on a PCIe drive. The document covers the Data Center Family and does not name the D5-P5336, so confirm the term for your drive with Solidigm. On a write-heavy workload the termination clause is the one to plan around: cover ends when Percentage Used reaches 100, whenever that happens. Our Enterprise SSD Percentage Used warning page covers what that reading means.

Solidigm and Wiwynn describe one approach, CSAL: a software layer that absorbs writes smaller than the indirection unit and writes large sequential data to the QLC drive, which they say keeps write amplification very close to 1. It is a platform change, not a drive setting.