A Higher Sequential Speed Means a Faster PC

A drive with a higher sequential speed does tend to score higher, but the number can’t tell you which drive will be faster in your PC. On my bench, 29 drives that all read at 7,000MB/s or more scored anywhere from 2,440 to 4,606 in PCMark 10 Storage, and across the 64 PCIe 4.0 drives sequential read speed explains about a third of the differences in score. I don’t have stopwatch times for real applications, only replayed storage traces, so this page tests the claim as far as a storage benchmark can. It is part of Flash Myths, Tested, and the Gen5 game loading myth uses the same trace approach on a faster drive generation.

Where the myth comes from

Sequential speed is the headline number on every SSD box, and it is real. A large file copy is a sequential job, and Solidigm’s traces of a game level load and of a Total War: Warhammer III battle load found mostly sequential reads, with a most common transfer size of 128KB in the level load. So a drive with a bigger number does move large files faster.

Going from there to a faster PC is where it goes wrong. Solidigm’s white paper on real-world PC storage points out that the numbers on a retail box are almost always measured at a very high 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., with 32 common, while the vast majority of real I/O it traced came at queue depth 1 or 2. Queue depth is how many requests are waiting on the drive at once. Across the 12 workloads it traced, 92% of all I/O operations were at queue depth 1 or 2.

Intel’s own evaluation guide for its 760p SSD makes the same distinction between kinds of benchmark. It marks a synthetic benchmark as not representing real-world use cases, and a trace-based one as possibly doing so.

What I measured

Both numbers come from the same retail drives on my test bench. One is CrystalDiskMark’s sequential read of 1MiB blocks at queue depth 8 (QD8), the median of repeated runs. That is a deep-queue number, like the ones a box prints. The other is the PCMark 10 Storage score from its Full System Drive test, the median of 3 runs. UL’s PCMark 10 technical guide says that test replays 23 traces of real applications and tasks, such as booting Windows 10 and starting Adobe programs, with 3 passes of each trace. Its score is the geometric mean of the bandwidth the traces reach and the inverse of their average access time.

The data covers the 84 retail NAND drives that have both results, 64 on PCIe 4.0 and 20 on PCIe 3.0. They’re the drives I tested for reviews, so they aren’t a random sample of every drive, and everything below comes from my test archive.

Figure 2 plots all 84. There is a trend. The correlation between sequential read speed and the score is 0.72, so faster-reading drives do score higher on average. Squared, that is 52%, so read speed accounts for about half of the differences in score, and the other half comes from something else. Part of the 0.72 comes from the two generations sitting in different places on both axes. Across the 64 PCIe 4.0 drives alone the correlation falls to 0.59, which is 35%.

Drives that read at the same speed

Figure 2’s right-hand band is the sharpest test. Twenty-nine PCIe 4.0 drives read at 7,000MB/s or more, all within 159MB/s of each other. Their scores run from 2,440 to 4,606.

Figure 3 sorts those 29 by score. The best is 1.9 times the lowest on a read speed that differs by about 2%. It shows up across the whole set too. Of the 84 drives, 1,047 pairs read within 5% of each other, and in 83 of those pairs the scores differ by 1.5 times or more. The widest pair differs by 2.4 times.

It happens in reverse too. In 11 of the 464 pairs where one drive reads at least twice as fast as the other, the faster-reading drive scored lower. Five drives in the table show the widest pairs and one reversal.

The first two rows are the widest pair within 5% on read speed. Next come the top and bottom of the 7,000MB/s group. Intel’s 670P is a PCIe 3.0 drive that reads at less than half the speed of the Patriot and still scores higher. I’m showing what the data records. These results don’t isolate a cause, so I make no claim about one.

What moves with the score isn’t the number on the box. Rows one and two differ by 4% at a queue depth of 8 and by 70% at a queue depth of 1, and rows three and four by under 1% and 39%. Sequential read at a queue depth of 1 has a correlation of 0.84 with the score, against 0.72 at a queue depth of 8, and random 4KB read at a queue depth of 1 reaches 0.77.

What PCIe 4.0 bought

Figure 4 compares the median of each generation. The median PCIe 4.0 drive reads 99% faster than the median PCIe 3.0 drive at a queue depth of 8, 6,979MB/s against 3,514MB/s. Its PCMark 10 score is 53% higher, 3,174 against 2,068. Results at QD1 sit nearer the score. Sequential read is 50% faster and random 4KB read 35% faster. Deep-queue speed nearly doubled, and the score rose by about half as much.

Solidigm’s own trace of the PCMark 10 Quick System Drive test found a mean queue depth of 1.2, with 96% of operations at queue depth 1 or 2, and concluded the test is useful for gauging how a drive does on real applications. It adds that PCMark 10 speeds up the replayed I/O, which stresses storage more than the original tasks did. I ran the longer Full System Drive test rather than the Quick one, so that trace shows how this style of test behaves, not what my runs did. Across my 84 drives the median PCMark 10 run moved 457MB/s, against a median sequential read of 5,225MB/s at QD8.

Why the box number misses

A deep-queue number shows how much a drive can do when the host keeps it busy, and a PC usually does not. A drive gets its speed by working on many commands at once across its channels and dies, and Flash Foundations Part 5 covers why a deep queue helps and why at a shallow one a lot of that hardware sits idle. On PCIe 4.0 the deep-queue read bunches up near 7,000MB/s, and 29 drives land within 2% of each other on it. At QD1 the same drives separate.

Beyond the drive there is the rest of the PC. Solidigm notes that if the bottleneck is elsewhere, such as a CPU under full load or exhausted memory, better storage won’t give a better experience. Its trace of the Warhammer III load showed the GPU at full utilization, and it judged the network the likeliest limit in joining a Zoom call, ahead of storage. That bridge from a score to a faster PC is the part I can’t measure.

What to do with this

Do not pay extra for a higher sequential number between two drives in the same class. Among the PCIe 4.0 drives here that read within 159MB/s of each other, the box number told me nothing about the score.

Look for numbers taken at a queue depth of 1: random 4KB read, sequential read at a queue depth of 1, and a trace-based score such as PCMark 10 Storage. Those moved with the score on my bench. If a review reports only deep-queue sequential results, ask what else it measured.

Sequential speed still matters for large transfers, such as copying a video project or cloning a drive. For those jobs, the box speed myth looks at the speed after the cache fills. For the method behind these tests, SSD Benchmarking 101 walks through queue depth, workload mix and why a spec sheet is a best case.