The Speed on the Box Is the Speed You Get

The write speed printed on an SSD’s box is real. My bench came within 10% of the rating on all seven drives I checked against their makers’ ratings. It is also the speed of a short write to an empty drive, and on most drives it lasts for less than a minute of continuous writing. After that, the same drive writes at a fraction of it. This is the first of the speed myths in Flash Myths, Tested, and the pages after it build on the cache cliff measured here. They ask whether a full cache stays slow and whether a higher sequential speed makes a faster PC. This page asks how long the number on the box holds, and what a drive writes at once it stops.

Where the myth comes from

Makers measure the write speed they print with a short test on a fresh drive, and Seagate, Kingston and Crucial say so in the small print. Seagate’s FireCuda 530 datasheet calls its figures fresh-out-of-box performance on a newly formatted drive, measured with CrystalDiskMarkCrystalDiskMark is a simple disk benchmark program for Windows that measures sequential and random read and write performance.. Western Digital rates the WD Black SN770 with CrystalDiskMark over a 1,000MB range. Samsung measures the 990 Pro with IOmeter while its TurboWrite cache is active, and its 970 EVO Plus datasheet adds that TurboWrite only works within a specific transfer size.

Those are short tests. My own CrystalDiskMark runs used its default 1GiB test file for the sequential write test (SEQ1M Q8T1, the median of the runs), and the smallest cache I measured held 13GB, so the whole test fits in the cache. They agree with the first seconds of my sustained-write test. On 51 of 64 drives, the burst my write test measured was within 5% of the drive’s CrystalDiskMark sequential write. The Samsung 990 EVO 2TB has no CrystalDiskMark data.

Figure 1 puts seven drives’ ratings, from six makers, beside my measurements. Three more ratings come from Kingston’s for the FURY Renegade, TEAMGROUP’s for the T-FORCE CARDEA Z44Q, and Crucial’s for the P3. Every rating is within 10% of the burst I measured. Furthest off is the 990 Pro, at 6,221MB/s in my write test against a 6,900MB/s rating, though its CrystalDiskMark write reached 6,805MB/s. Those after-cache dots are the part the box leaves out. They run from 163MB/s on the Crucial P3 2TB, about 5% of its rating, to 3,163MB/s on the FireCuda 530 2TB, about 46% of its rating.

Of the ratings above, only Samsung’s gives a number for after the cache. It rates the 970 EVO Plus 2TB at 1,750MB/s once the TurboWrite region is used up. My bench measured 1,697MB/s.

What my bench shows

Every drive here ran the same write-saturation test, which I describe in how I made SSD validation repeatable. It writes sequential 1MiB blocks at a queue depthIOPS counts commands finished per second. Queue depth is how many commands are outstanding at once, and a drive needs a deeper queue to reach its top IOPS. of 32, straight to the raw disk, starting empty. Runs lasted between 15 minutes and about two hours depending on the drive. That was long enough to write past the cache on 64 of the 65 drives.

Figure 2 shows three drives with three different answers. Here the FireCuda 530 2TB holds about 6,900MB/s for its first 226GB, then settles between about 3,800 and 3,900MB/s for the rest of the plotted range, and drops lower later in the run. Over almost 6TB, its average after the cache was 3,163MB/s. Every after-cache average on this page runs from the cliff to the end of that drive’s run, so it depends partly on how long the run was. Likewise the 970 EVO Plus 2TB holds about 3,100MB/s for only 84GB, then settles near 1,700MB/s. The WD Black SN770 2TB has by far the biggest cache of the three, writing between 4,500 and 4,800MB/s for 714GB, and then falls to about 420MB/s.

Across the whole set, 64 of the 65 retail drives fell off a cliff like these. The PNY LX3030 2TB was the one that did not. It held about 3,000MB/s for its whole 15-minute run. It wrote 2,790GB in that time, more than its own capacity, so no cache limit ever showed.

Figure 3 plots every drive’s burst against its average after the cache. Half of the drives kept about 26% or less of their burst speed once the cache was full, and the range ran from 4% to 59%. After the cache, 23 drives averaged under 1,000MB/s and 11 averaged under 500MB/s. That burst also says little about which drive will hold up. The 21 drives that burst at 6,000MB/s or more averaged anywhere from 990MB/s to 3,299MB/s after the cache. That is a threefold spread between drives with similar bursts.

Why the speed drops

On most drives, the cache is what the box number measures. A TLC or QLC drive writes incoming data to part of its NAND in SLC mode first, which stores one bit per cell and programs much faster than the cell’s native mode. It moves that data to its permanent home later. Flash Foundations Part 9 explains the cache, and Part 6 explains the move, called folding. Once the SLC area is full, new writes go to the cells in their native mode, sometimes while the drive is still folding the cached data, so the speed after the cliff is likely the native write speed or less. That is the lower line in each curve above. Samsung’s 850 EVO datasheet, for an older SATA drive, describes the same exit for its TurboWrite buffer. A continuous write that outgrows the buffer is processed at slower After TurboWrite speeds, and TurboWrite performance resumes once the buffer is cleared.

The interface limits the burst and says much less about what comes after it. Figure 4 compares the medians. PCIe 4.0 drives burst at about 5,540MB/s against about 3,050MB/s for the PCIe 3.0 drives, 82% faster. After the cache, the PCIe 4.0 median is 1,352MB/s against 1,007MB/s, only 34% faster. Newer drives earn their bigger box numbers in the first minute, and much less of that lead survives a long write.

Every run here started on an empty drive, which is the best case. A drive with a dynamic cache builds it from free space, so the cache shrinks as the drive fills. Part 9 covers how a full drive behaves. Cache sizes below are the most each drive offered, not what it offers half full.

What to do with this

For most desktop use, the box number is the speed you see. Saving files, installing a program and copying up to about 30GB at once all finish inside the cache on 62 of the 64 drives that hit a cliff.

Figure 5 shows how much each drive wrote at burst speed before it dropped. The median drive wrote 214GB, and 48 of the 64 wrote at least 100GB. Extremes were 13GB and 1,046GB. On time, half the caches filled within 39 seconds of continuous writing, 39 of 64 within a minute and 58 of 64 within two minutes.

If you move hundreds of GB at a time, such as video projects, backups or a drive clone, the box number will not tell you what you get. Look for the after-cache speed in a review that writes past the cache, because two drives with similar bursts can differ threefold there. A larger capacity buys a larger cache. In all seven models I tested at two capacities, the bigger drive had the bigger cache, about double in most. It does not always buy a faster drive after the cache, though. Two of those seven were clearly slower after the cache at the larger capacity.

Whether a full cache stays slow is the next question, and the next page takes it up with the same bench.