Slow NVMe performance on DIMM2 slot. Slower than laptop.

[Saturn7]

Crystal Bench with 601 BIOS



Crystal Bench from my Laptop with NVMe drive



I have reset the bios and cleared all overclock settings.

This is causing the system to run slower than my laptop. Loading up photoshop feels like I have a sata platter drive.

Random 4k Q32T1 benchmark is the problem.
IOPS are showing about 24k to 30k should be benching 200k iops minimum.

I can't use the main DIMM slot on the motherboard to test it right now because I have hardtubing for watercooling and can't remove my graphics card without draining the whole system and undoing the fittings which would take a while.

Can anybody confirm if they are getting better results with crystal mark?

-----------------------------------------------------------------------
CrystalDiskMark 5.2.2 x64 (C) 2007-2017 hiyohiyo
Crystal Dew World : http://crystalmark.info/
-----------------------------------------------------------------------
* MB/s = 1,000,000 bytes/s [SATA/600 = 600,000,000 bytes/s]
* KB = 1000 bytes, KiB = 1024 bytes

Sequential Read (Q= 32,T= 1) : 3267.551 MB/s
Sequential Write (Q= 32,T= 1) : 2131.028 MB/s
Random Read 4KiB (Q= 32,T= 1) : 126.813 MB/s [ 30960.2 IOPS]
Random Write 4KiB (Q= 32,T= 1) : 102.182 MB/s [ 24946.8 IOPS]
Sequential Read (T= 1) : 2146.020 MB/s
Sequential Write (T= 1) : 2031.733 MB/s
Random Read 4KiB (Q= 1,T= 1) : 37.151 MB/s [ 9070.1 IOPS]
Random Write 4KiB (Q= 1,T= 1) : 78.816 MB/s [ 19242.2 IOPS]

Test : 1024 MiB [C: 18.9% (180.4/953.3 GiB)] (x5) [Interval=5 sec]
Date : 2017/09/19 21:16:19
OS : Windows 10 Professional [10.0 Build 15063] (x64)

[ROGFanboy]

Korth wrote:
Again, different hardware provides different results. That benchmark was run on a 1TB Samsung 960 NVMe and an ASRock Z270 Extreme4 motherboard. Not an ASUS X399 Zenith Extreme motherboard. Different chipset, different architecture, different design, different implementation, different priorities, different performances, different tradeoffs, different bottlenecks.

Also remember that CrystalDiskMark is a synthetic measure. It's not profoundly meaningful (accurate) in itself, it's only meaningful as a reference comparable to other measures.

So you're telling me that a 600 euro motherboard cuts NVMe performance in half? With all due respect, but that makes absolutely no sense at all. I expect nothing but top performance, especially from a chip and motherboard like this. I am willing to compromise some performance, sure, but not over 50%. Outrageous.

Is there any other way I can truly test the performance of my 960 Evo? I figured CrystalDiskMark is a solid indication, but if you say it's not, then it's probably a good idea to test it in a different way? 🙂

Thanks!

EDIT: I installed Samsung NVM Express Driver 2.2 and got a performance improvement. Still not even remotely close to what it's supposed to be.



EDIT 2: Sorry! The performance increase was probably just a coincidence. The default Windows drivers are still installed somehow.



I am completely lost here.

[Ljugtomten]

ROGFanboy wrote:
So you're telling me that a 600 euro motherboard cuts NVMe performance in half? With all due respect, but that makes absolutely no sense at all. I expect nothing but top performance, especially from a chip and motherboard like this. I am willing to compromise some performance, sure, but not over 50%. Outrageous.

Is there any other way I can truly test the performance of my 960 Evo? I figured CrystalDiskMark is a solid indication, but if you say it's not, then it's probably a good idea to test it in a different way? 🙂

Thanks!

EDIT: I installed Samsung NVM Express Driver 2.2 and got a performance improvement. Still not even remotely close to what it's supposed to be.



EDIT 2: Sorry! The performance increase was probably just a coincidence. The default Windows drivers are still installed somehow.



I am completely lost here.

Have a look in Device Manager in the category for Storage Controllers and not the individual disks and you should see "Samsung NVMe Controller".
It is listed twice for me, have a 960 Pro and a 960 Evo.

[ROGFanboy]

Ljugtomten wrote:
Have a look in Device Manager in the category for Storage Controllers and not the individual disks and you should see "Samsung NVMe Controller".
It is listed twice for me, have a 960 Pro and a 960 Evo.

You're right, it's there. Explains the small performance gains. Thanks! Unfortunately the problem is still not fixed. Maybe it's something that ASUS has to patch with an update?

[relm56]

In my case it seems to be due to throttling. The 960 evo is hitting upwards of 80 C when I run CrystalDiskMark. I'll try adding a fan there and see how it performs then.

[Gogoltron]

i got the same problem with asus prime X399 and Samsung 960 EVO 1000GB very slow. And all computer on windows 10 seems to be very slow
the cpu is a RYZEN 1950X with 32gb
please help
thank you

f

[relm56]

IvoSilva wrote:
Your EVO seems to be the boot/OS drive (C:\) Then it would be normal for it not to achieve it's maximum performance since it it serving as the OS drive and will never achieve its peak performance because of it.

Anyone wanting the nvme max performance must use it as a standalone drive unrelated to OS and make sure it is cooled appropriately in order to avoid thermal throttling.

It is the boot drive yes. I'd expect some overhead due to that but compared to theirs, I'm getting 35% less on the 4k Q32T1 portion, seems more of a difference than it should be.

Maybe it is throttling however - I'll check temps.

[Saturn7]

Korth wrote:
Actually looks like not just two different SSD drives, but two entirely different SSD models, based off those numbers.

Which TR4 proc and how much RAM on that X399?
Which laptop model (which chipset, CPU, RAM)?
What other hardware (and other drives) plugged into those systems?
What other software is running (why is desktop consuming 19% resources vs laptop consuming 7% resources)?
Are you running any drive-specific firmware or software like RAM Cache, RAM Disk, RAPID/Magician?

I wonder if (Windows) chipset drivers can produce such significant differences. People say the AMD drivers are still imperfect but those numbers (if on the same SSD hardware) just seem too divergent to blame entirely on drivers.

System: 1950x with Corsiar Vengance 3444 Mhz ram running at stock 2133Mhz. (everything running at stock, but CPU easily overclocks to 4 Ghz and RAM can go to 3000Mhz.)

Yes they are different model drives but both samsung 1TB nvme dirves, the laptop was just for reference. Just to show that it's IOPS are much higher than the 1TB Samsung 960 Pro, M.2 PCIe NVMe which is currently in the X399 system. Laptop drive is slower by specs, but benchmark shows it's nearly 5 times faster for random read and writes.

ASUS AI software uses lots of CPU don't know why. Otherwise it's a fresh system install.

I did install Magician afterwards to update the firmware of the Samsung drive. But results are basically the same.

[Korth]

"Yes they are different model drives". Answers your question right there. You'd get essentially identical benchmarks - though reversed - if you swapped the NVMe drives between your machines.

Just because they're both made by Samsung and both 1TB capacity and even both NVMe-capable doesn't mean they're the same machine. Samsung makes lots of different (and always changing, always improving) 1TB SSDs.

A Ford Mustang and a Ford Escort are different machines with different performances. Even though both are made by Ford and both made to be driven on the same roads with the same speed limits.

[Saturn7]

Korth wrote:
"Yes they are different model drives". Answers your question right there. You'd get essentially identical benchmarks - though reversed - if you swapped the NVMe drives between your machines.

Just because they're both made by Samsung and both 1TB capacity and even both NVMe-capable doesn't mean they're the same machine. Samsung makes lots of different (and always changing, always improving) 1TB SSDs.

A Ford Mustang and a Ford Escort are different machines with different performances. Even though both are made by Ford and both made to be driven on the same roads with the same speed limits.

The lower spec drive should not out perform the high spec drive by 5 to 1. It's just for comparison to show there is a performance problem.

In the same way if that if a Ford Escort is faster than a million dollar Bugatti there clearly is a problem with the Bugatti. I don't think it's that hard to understand.

Forget about the laptop. The drive in the system is rated at 300k IOPS and it's only getting 24k

[dejanh]

I'm going to necro this thread because I have been struggling and ignoring NVMe random 4k read and write issues for years, and have finally had it. I have two Samsung 980 Pros 512GB each and a single 970 Pro Plus 2TB drive and I am getting ridiculously low random 4k R/W at QD32. Like when I say ridiculous, I mean I get 90% lower performance than spec. My 2TB 970 Evo Plus running connected to the motherboard slot directly (not the DIMM.2 card) gets 360k IOPS random read, and only 20k IOPS random write. Even my ROG X13 Flow laptop smokes this using the no-name drive it comes with. The write IOPS on the laptop are literally 5x better.

It does not matter if the drive is OS drive or storage drive, and it does not matter whether Samsung NVMe driver is installed or not. The performance is essentially the same abysmal performance. I am so fed up with this that I am about to make a trip to the store to buy a new Ryzen machine instead of the Threadripper 2950x setup I currently have just to prove the point that something is really messed up with this chipset and NVMe. Has anyone been able to figure this out? Note I am now running Windows 11, I used to run Windows 10, and it doesn't matter whether it is a fresh install or not. Performance is abysmal either way. The 2950x is running on a Zenith Extreme Alpha board, running the latest UEFI/BIOS.