Can you trust a 512 MB benchmark? We checked every DiskBoard result for cache effects

Drafted from every result uploaded to DiskBoard since November 2023.

DiskBoard's default test writes and reads a 512 MB file. A reasonable objection: 512 MB fits comfortably in RAM, so is the read test measuring the drive or the operating system's file cache? We went through every uploaded result to find out. Short version: about three percent of results are inflated by a cache, the cache is usually not the operating system's, and you can spot the bad ones by eye.

How to catch a cached result

A drive cannot be faster than the wire it hangs on. SATA tops out around 560 MB/s. PCIe 5 x4, the fastest NVMe slot today, tops out around 14,500 MB/s. Any result above those numbers did not come from the drive.

There is a second tell. Real SSDs read somewhat faster than they write, typically 1.1 to 2 times. A RAM cache makes reads five, ten or twenty times faster than writes. So we also flag any result where read is more than three times write.

What the data shows

SATA: 51 of 1,831 sequential results (2.8 percent) read faster than SATA physically allows. Forty-one of them read above 1,000 MB/s, and the top of the list reads at 12,449 MB/s. That is not a slightly optimistic number; it is twenty times the bus.

NVMe: 39 of 5,651 results read above 10,000 MB/s and 3 above 15,000. The record is 25,653 MB/s on a Crucial P310 1TB, which is a PCIe 4 drive rated for 7,000. That run used 12 threads on a 16 GB machine.

Overall: applying both tests removes 4.4 percent of NVMe results, 7.6 percent of SATA and 9.1 percent of USB. The higher USB and SATA figures are partly false positives: shingled hard drives genuinely write three or four times slower than they read, and the ratio test catches them too. When we look only at the physical-limit test, the true cache rate is about 3 percent.

It is mostly not the operating system

If Windows or macOS were caching the 512 MB file, two things would follow. Larger test files would be cached less, and machines with more RAM would be cached more. Neither happens.

Doubling the file changes nothing. And the flagged share falls as RAM rises: 13 percent on machines with 8 GB or less, 4 percent at 32 GB, 3 percent at 64 GB and above. The low-RAM number is the SMR hard drives again, not caching.

What the inflated SATA results have in common is the drive. Every result above 10,000 MB/s on SATA is a Samsung 860 EVO or 860 QVO, on a machine with 32 or 64 GB of RAM, and each one also writes at 5,000 to 9,000 MB/s, which SATA cannot do either. That is the signature of Samsung Magician's RAPID mode, which uses system RAM as a write-back cache in front of the drive. Crucial's Storage Executive has the same feature, called Momentum Cache, and it is the likeliest explanation for the one-in-ten Crucial P3 results that exceed the PCIe 3 limit.

DiskBoard's test does not appear to be fooled by the ordinary operating system cache. It is fooled, like every benchmark, by a vendor utility that has inserted a RAM cache between the application and the disk. From the application's point of view, that cache is the disk.

What we do about it

The rankings on DiskBoard will exclude results that fail the physical-limit test, and show the sample count next to every number so you can judge how much to trust it. We are not going to remove the ratio-flagged results wholesale, because some of them are real: a shingled hard drive that reads at 180 MB/s and writes at 40 is telling you something true about that drive.

What you can do

• If you want to know what your drive does, turn off RAPID mode or Momentum Cache before running any benchmark, then turn it back on if you like it. Those utilities make everyday use feel faster; they just do not measure the drive.
• Run with one thread if you want a number you can compare with other people's. Most uploads are single-threaded. More threads show what an NVMe drive can do under load, and that is a different, also useful, question.
• The 512 MB default is fine. The data says a 1 GB file gives the same answer.

Method

Sequential results only. "Physical limit" is 600 MB/s for SATA and ATA, 15,000 MB/s for NVMe, 4,000 MB/s for USB (to allow for USB4 and Thunderbolt enclosures), 320 MB/s for SD. Ratio test is read divided by write greater than 3. Percentages are of sequential uploads at the time of writing.