Question Bad SSD & NVMe performance after switching from Haswell to AMD Ryzen 3900X

Aug 16, 2019
3
0
10
Hi guys,

I am very happy with my upgrade to a Ryzen 3900X from an i5 4690k so far, setup was easy and performance is great - except for my SSD & NVMe drives.

Now I know that UserBenchmark has its flaws and is biased towards Intel, but for drive benchmarks it's usually pretty accurate.
On my old system, all drives performed well within the top 30% percentile. On the new one, they are performing below expectations, in the low 70% percentile. It's not a space issue, since they are even less full on the new System. Sequential Speeds are fine (but those don't really matter to me) but 4K is where the problem lies.
See for yourself:
Especially Deep Queue 4K seems to be the culprit (which is sad since 4K actually corresponds to real-world performance the most). This is not specific to UserBenchmark, AS SSD and CrystalDiskMark also report lower 4K scores.
When checking other benchmarks with exactly the same board and CPU, they get great scores on their SSD. What could I try to fix it? Doesn't seem to be a hardware issue.

Build:
  • Ryzen 3900X
  • Gigabyte Aorus Elite X570
  • G-Skill Flare-X DDR4 3200 CL14
  • Samsung 960 Evo, 860 Evo and 850 Evo
What I've tried:
  • CMOS Reset
  • Disabling HPET and C-State
  • Using Windows NVMe driver instead of Samsung
  • Reinstalling newest AMD chipset driver
  • Update to latest beta BIOS
  • Using chipset Driver from board manufacturer website
Also, TRIM and write caching are enabled and all of the drives are on the newest firmware.
Any idea what else I could try? Since other users with that exact board are getting good scores, I really have no idea what I am doing wrong.
 
Aug 16, 2019
3
0
10
High queue depth greater than 4 almost never happens in real world. Don't obsess about Q32 performance.
Generally I'd agree, although this setup is meant to be used as a workstation with virtual machines and intensive database I/O. (Hence I went for the 3900X). 32 is not realistic, sure, but I do have some serious parallelization going on.

Apart from that, regular 1Q4K performance got worse by around 30% as well. This is definitely something you can notice in regular single user desktop operation.

Generally, I'd like to know how upgrading from an obsolete chipset/processor could have such an impact. Is Intel just doing I/O better?