An SSD data sheet prints its biggest numbers with the most conditions attached. Samsung’s 990 PRO data sheet rates the 2TB model at up to 7,450MB/s reads, 1,200TBW (terabytes written), and an MTBF (mean time between failures) of 1.5 million hours, under a five-year limited warranty. Six numbered notes at the foot of the page say what those figures assume. Part 3 read the interface and link numbers on the box, and this part reads the rest of the sheet.
Storage and Memory in Practice
Storage and Memory in Practice, Part 4: How to Read an SSD Spec Sheet
What each line on an SSD data sheet means: the conditions behind rated speeds, TBW and DWPD as one write budget, MTBF as a fleet rate, and warranty terms.
On this page 13 sections
To see which lines compare across makers, I’ll set the 990 PRO sheet beside two others. Kingston’s NV3 is a budget PCIe 4.0 drive, and Micron’s 7450 is built for data centers. All three makers state the same kinds of numbers under different conditions, and a few of those numbers turn into something about your own use once the arithmetic is done. Most figures here are from a maker’s sheet or from my own drives, and the rest name their tester or study.
What a rated speed assumes
Every speed on a consumer sheet is a best case, and its notes say which one. Samsung heads the 990 PRO’s performance table “Up to” and measured it with IOmeter 1.1.0 on an AMD Ryzen 7 5800X system with the drive as a secondary drive. Its write figures were taken with Intelligent TurboWrite active. TurboWrite is Samsung’s SLC cache, and the note says it works only within a specific data transfer size without giving the size. Kingston bases the NV3’s speeds on out-of-box performance on a PCIe 4.0 motherboard.
What a rated speed assumes
Three sheets, three sets of test conditions
What each maker’s notes say about how its speed ratings were measured
Set side by side, only Samsung’s sheet names a tool and a test system, and only Micron’s rates a drive written to steady state.
Fresh or steady state
A data center sheet measures a different drive state. Micron rates the 7450 at steady state as SNIA’s Solid State Storage Performance Test Specification defines it, with 128KB transfers for its sequential figures and 4KB transfers for its random ones. A fresh drive writing into its cache and a drive written until its speed stops changing are two different measurements, which Flash Foundations Part 9 explains, so a consumer rating and an enterprise rating can’t be set side by side.
Queue depth and threads
Random ratings depend on how many requests are in flight, and a sheet states each one at a queue depth (QD) and a thread count. Samsung’s random read rating for the 990 PRO 2TB at QD32 with 16 threads is about 64 times its rating at QD1, so a random figure means little until its queue depth is named. Part 3 gives both figures and what each measures.
Ratings against my bench
Those ratings hold for what they describe. In my write runs, every one of the seven drives I could match to a maker’s rating came within 10% of the burst I measured, and the 990 PRO 2TB wrote 6,221MB/s in its first burst against its 6,900MB/s rating. Past the cache the speed falls, and The Speed on the Box Is the Speed You Get shows by how much, drive by drive, with the method behind each run.
Each capacity is its own drive
A sheet that covers a product line gives each capacity its own row, and the rows differ. Kingston rates the NV3 at 5,000/3,000MB/s read/write at 500GB, 6,000/4,000MB/s at 1TB, and 6,000/5,000MB/s at 2TB and 4TB, so the 500GB model writes at 60% of the 2TB model’s rating. Its endurance doubles with each step, from 160TBW at 500GB to 1,280TBW at 4TB.
Each capacity is its own drive
One NV3 sheet, three different write ratings
Kingston NV3, M.2 2280: rated sequential read and write by capacity
Only the 2TB and 4TB bars match, and the write bars climb with capacity while the read bars level off from 1TB up.
Samsung’s 1TB and 2TB 990 PRO share their sequential ratings, and the rows still part further down. At 1TB the drive is rated at 1,200K random read IOPS against 1,400K, carries 1GB of LPDDR4 DRAM against 2GB, and is rated for 600TBW against 1,200TBW. Reviews of one capacity speak for another only where their rows match, and the controller behind each row can differ too, as the section on what the sheet leaves out shows.
TBW and DWPD are one write budget
Consumer sheets state endurance in TBW and data center sheets in DWPD (drive writes per day), and either converts to the other. SNIA’s endurance white paper sets out the conversion:
DWPD = TBW / (capacity × warranty days)
TBW and capacity are both in TB, and a five-year warranty is 1,825 days. For the 990 PRO 2TB that is 1,200 / (2 × 1,825) = 0.33 DWPD, so its warranty assumes a third of the drive written each day for five years. Kingston’s NV3 2TB comes to 640 / (2 × 1,825) = 0.18 DWPD.
Micron rates the 7450 PRO at 1 DWPD for 4KB random writes, and its 1,920GB model’s 3,650TB of them works out to 1.04. Its 7450 MAX is rated at 3 DWPD, and the 1,600GB model’s 8,700TB works out to 2.98.
TBW and DWPD are one write budget
The same five-year budget, stated as drive writes per day
DWPD = TBW / (capacity × 1,825 days)
Both data center drives budget more writes a day than either client drive, and the 7450 MAX budgets close to three full drive writes.
The workload behind the number
A TBW figure is only as good as the write pattern behind it. Kingston derives the NV3’s from JEDEC’s client workload. Micron’s table has two figures for each 7450 capacity, assuming the drive is full. For the 1,920GB PRO those are 3,650TB of 4KB random writes and 12,800TB of 128KB sequential writes, 3.5 times as much, and Micron adds that actual lifetime will vary by workload.
Samsung’s note says its endurance results comply with JEDEC’s endurance test standard, and a JEDEC presentation on JESD218 lists its classes. The client class assumes 40°C for 8 hours a day and data kept for one year at 30°C with the power off, and the enterprise class assumes 55°C around the clock and three months at 40°C. How a maker builds its rating from those conditions is in Flash Foundations Part 9, and the SSD endurance board ranks client drives by TBW per TB of capacity.
How long a TBW rating lasts
A TBW rating turns into years once you know your own write rate. My 4TB system drive had taken 42.9TB of host writes over 10,726 power-on hours, about 96GB for each day it was on, and it reported a Percentage Used of 1%. At that rate the 990 PRO 2TB’s 1,200TB would last 12,500 days, about 34 years, and the NV3 2TB’s 640TB about 18 years. Both would run out of warranty years long before they ran out of TBW. Do SSDs Wear Out in a Couple of Years? puts my system drive and my bench drives beside these ratings in more detail.
How long a TBW rating lasts
At a light write rate, the rating outlasts the warranty
Years to reach rated TBW against data written each day
A heavy writer changes the answer. Say a drive takes 800GB a day. In that case the 990 PRO 2TB would reach 1,200TB in about 4.1 years and the NV3 2TB would reach 640TB in about 2.2 years, both inside five years.
Each drive also keeps its own count. NVMe’s base specification defines Percentage Used in the SMART/Health log as the maker’s estimate of how much of the drive’s life is used, where 100 means the estimate is spent, and the value can go past 100. Kingston’s note says an NV3 at its warranty limit shows 100 or more. You can read the field, and the bytes written beside it, from a drive’s log with the NVMe SMART Decoder.
MTBF is a fleet rate, not a lifespan
An MTBF of 1.5 million hours reads like 171 years, and it means nothing of the kind. A mean time between failures is a rate taken across a large number of drives, and it doesn’t estimate how long one drive will last. Micron gives the 7450 a mean time to failure (MTTF) of 2 million hours instead, and says plainly that it is “based on population statistics not relevant to individual units.” Kingston rates the NV3 at 2 million hours too.
Under a constant failure rate, an MTBF converts to an annualized failure rate (AFR). AFR is the share of drives expected to fail in a year:
AFR = 1 − exp(−8,760 / MTBF)
Here 8,760 is the hours in a year, and MTBF is in hours. Samsung’s 1.5 million hours gives 0.58%, about six drives in every 1,000 running all year, and Kingston’s 2 million hours gives 0.44%, about four. For any other rating, the Reliability Math Lab runs the conversion.
MTBF is a fleet rate, not a lifespan
An MTBF is a yearly failure share across many drives
AFR = 1 − exp(−8,760 / MTBF), for a constant failure rate
Real fleets don’t fail at one flat rate. A study of the flash in Facebook’s data centers found that in every platform the 10% of drives with the most errors held over 80% of the uncorrectable errors, and that failure rates rose and fell with the data written instead of holding flat. Flash Foundations Part 9 covers the study. Between these two consumer drives the MTBF line is worth about one and a half failures per 1,000 drives a year, which is why it rarely decides between them.
Which limit ends the warranty
Warranty wording is where an endurance number becomes a promise, and it decides which limit ends coverage. Samsung covers the 990 PRO for “the stated time period or the TBW, whichever comes first,” five years limited. Sandisk words the Optimus GX 7100M warranty the same way, five years limited or the endurance limit. Kingston ties the NV3’s five years to Percentage Used instead, which its SSD Manager reads, and its sheet doesn’t say how that count maps to the 640TB figure.
Which limit ends the warranty
Each warranty ends on the calendar or on the writes
The two limits each maker’s warranty names
At my 96GB a day, each of these warranties ends on the calendar. At 800GB a day, the 990 PRO 2TB would use its 1,200TB in about 4.1 years, so for a heavy writer the TBW figure sets the warranty’s length. Each is a limited warranty, and Samsung’s sheet points to a separate warranty statement for the terms rather than printing them.
Kingston’s sheet adds that the NV3 is designed for desktop and notebook workloads and isn’t intended for server environments. That line is worth reading before a consumer drive goes into a NAS, which SSDs in a NAS covers.
What the sheet leaves out
Some lines a buyer wants aren’t on the sheet at all. Samsung’s names an in-house controller and its own V-NAND TLC. Kingston’s names no controller and lists the NV3’s NAND only as 3D. Neither has a write speed for after the cache fills. Samsung’s TurboWrite note says the cache works only within a specific data transfer size and sends readers to a service center for detail.
Testing fills some of the gaps. SSD-Tester found a Phison E27T controller in its NV3 1TB and a Silicon Motion SM2268XT2 in its NV3 2TB, both with QLC flash and no DRAM, so the two units it tested, from two capacities on one sheet, carried different controllers.
What the sheet leaves out
Two sheets, and what testing had to fill in
Parts each sheet names, beside what SSD-Tester found inside the NV3
On the NV3, only the testing column names the parts, and no column gives the write speed past the cache, which takes a long write test like the ones on the box-speed page.
Parts that change under one name
Parts can change after launch, too. Flash Foundations Part 9 covers a controller change on Samsung’s 970 EVO Plus that TechSpot reported. Crucial swapped TLC flash for QLC in the P2 without changing its model number. I retested the P2 for Tom’s Hardware after the swap, and the 500GB model was then rated for writes of up to 940MB/s, where the original review sample had written at over 1.8GB/s. The Field Practicalities covers that case, and Best SSDs names the controller, DRAM, and NAND behind each pick.
What to watch
When a new sheet or a new revision of one comes out, check the row for the capacity you’ll buy, turn its TBW into years at your own write rate, and read what ends the warranty. A note that changes the test tool or the cache state changes what a speed rating means, and a parts change can reach a drive without any new sheet at all. Per TB of capacity, the NV3’s 320TBW is about half the 990 PRO’s 600TBW, so on a budget drive the TBW line is the one to read most closely.
What to watch
Two 2TB sheets, compared line by line
Samsung 990 PRO 2TB against Kingston NV3 2TB, with this page’s arithmetic
Next in this series is HDD versus SSD on cost per TB, power, and rebuild time. Current picks by tier are in Best SSDs, and the SSD endurance board ranks client drives by rated TBW per TB.
Sources
Primary is the original paper, standard or datasheet. Secondary is a review, a vendor summary or a news report. Measured is my own test data.
Show all 13 sources, 9 primary
- PrimaryA Large-Scale Study of Flash Memory Failures in the FieldACM SIGMETRICS 2015, 2015
- PrimaryUnsavory Flash Swap: Re-Testing Crucial’s P2 SSD After QLC DowngradeTom's Hardware, 2021
- PrimaryNVM Express Base Specification, Revision 2.1NVM Express, Inc., 2024
- PrimaryMicron 7450 SSD Series Technical Product SpecificationMicron Technology, 2026
- PrimaryEndurance of NVMe, SAS, and SATA SSDsSNIA
- PrimaryJEDEC SSD Specifications Explained: JESD218 and JESD219 (Alvin Cox, Seagate)JEDEC
- PrimaryKingston NV3: data sheetKingston
- PrimarySamsung NVMe SSD 990 PRO Data Sheet, Revision 1.0Samsung Electronics
- PrimarySandisk Optimus GX 7100M NVMe SSD data sheetdocuments.sandisk.com
- SecondarySamsung is swapping parts in their 970 Evo Plus SSDs and sabotaging performance (TechSpot)TechSpot
- SecondaryTEST: Kingston NV3 1TB M.2 SSDSSD-Tester
- SecondaryTEST: Kingston NV3 2TB M.2 SSDSSD-Tester
- MeasuredThe author’s own measurements

















































































































