M2 Air 2TB SSD is 3.5GB/s, 7x slower on QD1 random write than 2018 MBP

[drecc]

M2 Air just arrived. AmorphousDiskMark 4.0 set to 5x 4GiB reports the following performance on the M2 Air (left) vs 2018 MBP 15" (right):

I wonder what could account for the huge difference in random 4K writes between the machines? 35 vs 272 is a huge difference.

 

[Gnattu]

272 is definitely with some caching mechanism and 35 looks about right to me for a TLC NAND base SSD when the SLC cache used up.

 

[drecc]

272 is definitely with some caching mechanism and 35 looks about right to me for a TLC NAND base SSD when the SLC cache used up.

Thanks, yes an SLC cache sounds like it must be the reason. Surely it can't be very expensive to put a few GB of SLC cache on a 2TB SSD. Even when I reduce the AmorphousDiskMark test size to 16MiB, there is not much of an improvement with random writes.

 

[Chancha]

Here are the SSD speed tests of a few of Macs we own. It does seem QD1 results are sometimes due to cache in play, or that the 5 pass 1GiB isn't enough to get it above the thershold of going in and out of the NAND. For that reason I tend to just look at the random write QD64 as a more meaningful benchmark for comparison.

That one 2018 13" MBP I have indeed has a not bad SSD in it, even beating out some 1TBs.

 

[drecc]

View attachment 2032560
Here are the SSD speed tests of a few of Macs we own. It does seem QD1 results are sometimes due to cache in play, or that the 5 pass 1GiB isn't enough to get it above the thershold of going in and out of the NAND. For that reason I tend to just look at the random write QD64 as a more meaningful benchmark for comparison.

That one 2018 13" MBP I have indeed has a not bad SSD in it, even beating out some 1TBs.

Thank you, that's very interesting - it makes me feel much better that even the M1 Max is not getting significantly higher random 4K writes than my M2 Air.

It looks like every single Intel Mac in those benchmarks scores massively higher than any M1 Mac.

 

[Gnattu]

Thank you, that's very interesting - it makes me feel much better that even the M1 Max is not getting significantly higher random 4K writes than my M2 Air.

It looks like every single Intel Mac in those benchmarks scores massively higher than any M1 Mac.

One thing to note is that starting T2, all SSD controllers are implemented by Apple(the T2 chip has an SSD controller), and we can see that the SEQ is improving but RND is having regression from T2 to M series. It seems like Apple now prefer not to use SLC cache for QD1 4K now.