Yes this is long. Short version is a 7 year old drive model can compete very well with a 1 year old model despite double the interface bandwidth, presumably because MLC NAND is itself faster. I’m not a storage expert, though.
I’ve had a 512GB Samsung 960 Pro for a long time and wanted to upgrade to a faster drive to take advantage of the full PCIe 4.0 connection on my motherboard and newer flash and controllers and whatnot, but I was waiting until I could get something really good instead of just good enough, because the 960 Pro was already good enough. I finally got a few bucks extra so I got an SK Hynix Platinum P41 which is rated for more than double the read speed and nearly triple the write speed.
I haven’t had a lot of time to “experience” it, but benchmarks were interesting and almost make me think I might not actually see any difference. I cloned the old drive and have been working on adjusting partitions since the new drive is 1TB. Boot times for Windows actually seem to be noticeably higher, oddly. I didn’t specifically time it before the replacement.
The numbers for the 960 Pro came from some months ago when I put in a secondary drive and decided to test them both. The 960 Pro was on the main M.2 slot connected to the CPU PCIe 4.0 (x4) lanes but the drive is only PCIe 3.0.
AS SSD Benchmark shows the anticipated faster speeds, though it wasn’t hitting quite the rated speeds. 2.34x sequential read and 2.92x sequential write speeds. 16% higher 4k read, but 16% lower 4k write speed. 4k-64thrd read was 29% higher with the P41, but write was 54% slower!
ATTO provides more detail and shows where the 960 Pro still wins hands down. At lower I/O sizes, the 960 Pro dominates the P41, nearly doubling both reads and writes for both throughput and IOPS. This is consistent up to 16KB, and becomes less overwhelming at 32KB, and only at 64KB does the P41 begin to outpace the 960 Pro. The P41 at that point continues to increase throughput as I/O size increases, but the 960 Pro never gets any faster. On the IOPS side, the 960 Pro plummets as the I/O size increases, while the P41 decreases much more slowly. By the end, the P41 is still doing more than 3x IOPS on write, and over 2x on reads.
So it seems to me that the MLC is SO fast that it probably could have still saturated a PCIe 4.0 connection even though it’s now 7 years old. The P41 should have been working in pseudo-SLC mode for it to reach the 6GBps speeds I was seeing, yet it still doesn’t perform as well as that old MLC at lower I/O sizes, which I bet is due to the limitations of the PCIe 3.0 connection and perhaps the older NVMe spec. (And the 960 Pro was even using Samsung’s custom NVMe driver.)
I’m going to run the benchmark again tomorrow, after the drive has had plenty of idle time to flush the SLC cache with all the data being copied and partitions moving around. I used GParted to move the C drive while it was still 630GB in size from the cloning, and it did it in bit-for-bit mode, so it literally wrote zeros to like 80% of the drive, despite moving it by only 300 MEGABYTES, which may have filled up the SLC cache causing the drive to write at native-TLC speeds at least partially. But given that it DID reach those high read and write speeds, it seems unlikely that it was significant, as reviews show the native-TLC speed is more like 1.5GBps.
If these tests indicate the true performance, then I can’t assume I’ll automatically get better performance simply due to it being a newer drive with a faster connection, despite it even being one of the highest-rated PCIe 4.0 drives (a year ago anyway). I’m not sure what my normal I/O size patterns look like. Most of my usage is web browsing these days (at least the stuff that impacts the main drive), plus playing Civilization 6 which isn’t demanding on storage. But these days, even 16KB would be considered a “tiny” file, so it may still be that the average I/O will be much higher and into the range where the P41 is faster.
It’s too bad that consumer SSDs will never have MLC again, and it’s pretty expensive for enterprise drives. I’d think with process shrinks it could still be useful in terms of capacity, but obviously not for really massive storage, with a reasonable price premium for the performance.
Oh none of these was actually near “full”, other than being partitioned (though I leave about 10% unpartitioned). I wanted a larger drive so I could consolidate the two I was using without having to shrink them, as sometimes I DO end up temporarily using a lot more space, so I could stick another larger drive in the second slot for another use. But it would only benchmark worse if it were so full that the pseudo-SLC cache available was reduced to the point that the benchmark run used it up, and then would get even worse if the benchmarks were writing enough data to use up unallocated blocks and prevent garbage collection before the next run, so that the NAND needed to perform the erase cycles during the tests. I’ve only actually got 354GB of data used out of the 932GB real space on the drive.
Interestingly, my first pass test this morning was actually worse than it was last night on the P41, however I had not forced TRIM to run and it was only about 5 hours of idle time in Windows. And I realized that BOTH runs in ATTO were actually slightly slower than the Inland (microcenter brand) PCIe3 drive in the secondary M.2 slot, at least up to the point that the PCIe3 bandwidth became a limitation.
I’ve got time to let things settle and do more testing, and see how the real usage feels, but it’s really starting to look like in my particular use case the extra cost for this drive may not be justified. Having higher max throughput on the occasional times that I’m writing a really large amount of very large files is nice, but I’m not sure it’s “twice the price” nice or that it happens often enough. In the past, I did a lot more stuff that could have made it justifiable, and I’ve still got that mindset when I look at parts.
Seeing that the random performance, which is the bulk of usage, is actually WORSE than even a cheap drive or my old main drive, is quite disappointing. I just hadn’t really considered the possibility that such an old drive could have been so far ahead of the pack back then that it remains a good choice today.