Low 4K scores for Taichi & 960 EVO

Here is what HWiNFO64 says about 960 EVO:
Warning Temperature Threshold:                                                  77 °C
Critical Temperature Threshold:                                                 79 °C

Latest official FW for the EVO, standard Win 10 NVMe driver.

I got an EKWB M2 rad and decent airflow around my 960. Temps during tests were at 35 deg C.

Here are the results:





What am I missing? Writes are better than reads? Since when?

Also, did anyone check thermals during the tests? We all know that NVMe devices are really prone to thermal throttling. 4K random, being last of the read tests, mean that the drive itself has been warmed up in advance, hence possible throttling.

In case of 4K Q32T1 settings, R/W performance will be under 400MB/s but if you change the settings to 4K Q32"T4", you will be able to get over 1,000MB/s @Ryzen environment.

Edited by CHO - 01 Sep 2017 at 9:08pm

Here's my 500GB 960 EVO in my ASRock X370 Killer SLI/AC board:



Looking at a benchmark result does not begin to tell the whole story of how it was achieved. Yes, storage drive bench marking has its tricks and tweaks to get better results. We can get some clues regarding what some of the situation is by looking closely at each of the screenshots in this thread.

First, if the drive being tested is the OS/Windows C: drive, the results will be lower than a non-OS drive. The benchmark program does not own the PC when it's running, so any IO done by Windows will interrupt the benchmark. Any non-OS drive will not have the benchmark interrupted by other IO.

The second and third Crystal screenshots in the first post that are supposed to be from a Gigabyte X370 board, show the drive being tested were the D: and H: drives. Both not the C: OS drive, so the results will be better. The first screenshot has the drive information cut off, so no idea.

All the other Crystal screenshots from ASRock boards in this thread are of the C:/OS drive. So not an identical comparison.

Second, the 250GB EVO shown (third Gigabyte screenshot in the first post) which is overall exceptional (with a 50MB/s 4K read) shows 0/233GiB used. That is an empty, new SSD, which will always have great performance compared to a SSD that has been used. Empty is more important than new, but new also helps provide a better result.

Next the things we cannot see: CPU OC, Memory OC, CPU power saving option settings, Windows Power Plan option settings, and Windows write cache buffer flushing setting. Do we know about any of these things in the superior benchmark results? No.

How did I get the 60MB/s 4K Read speed? I disabled Cool n Quiet and C State support in the UEFI. Even though IO is not CPU intensive, the latency caused by the processor cores switching in and out of the high (C6) C State can be seen in benchmark results. I know it shows up with Intel systems and now I know the same is true for AMD Ryzen. Not a surprise.

Gigabyte is (in)famous for setting their default UEFI/BIOS options to provide better performance when their mother boards are tested in reviews, with all default settings. Meaning CPU power saving options on Auto is really Disabled. My ASRock X370 board's UEFI and Cool and Quiet, C6 and Global C6 Support Enabled by default.

I'm using the AMD Ryzen Balanced Power Plan, but this applies to any power plan. PCI Express Link State Power Management is set to Off. The Ultra M.2 slots use the PCIe lanes from the CPU, so any power saving on those will cause latency and slower speeds.

I haven't tried the High Performance power plan, which I know gives the best benchmark results with Intel systems. I will say in general the 4K performance of any SSD seems to be lower with my Ryzen system than my Intel systems. Of course, Intel has had PCIe 3.0 M.2 support since their Z97 boards released in 2014. We are using the first AMD Ryzen NVMe Option ROMs in our board's UEFIs, which should be improved down the road.

Here's one fully tricked benchmark of a 960 EVO on an Intel Z270 board:

also there doesn't seem to be a problem in this test done on a gigabyte board