very slow nvme speed compared to my friend who has the same drive

[seccentral]

Can you test your nvme in friends mobo and if performs same you will know that drive has issues if not then it's rest of your PC.

doing that today

Ask your buddy to let you borrow his Drive to see if it does the same thing in your system.

From everything I read even if the SLC cache fills up you should still be at HDD speeds around 100MBs.

You also should try copying over the exact same game to verify speeds last I checked guild wars 2 used 1000s of tiny files even some small patches being 12k files.

doing that today, gw2 has one big file, maybe guildwars 1 had many small files, but i dont know that one.

This.

Have you got chipset drivers installed try the same game not just a "similar" sized game, both drives copying and writing to need to be the same for read/write speeds otherwise you may as well compare apples to oranges and scratch your head why you've ended up with a banana

I do have the drivers installed https://drivers.amd.com/drivers/amd-chipset-drivers_18.50.0422.exe they are the latest. will test that today since we both live 2 minutes away from each other.

He has a thread open on the Asrock forums also saying its a 600P not the 660P but on there he claims both systems are his.

Found the thread googling if the problem was widespread on his mobo.

yes i do. and we basically live across the street from each other, im the one who assembled his pc too so for simplicity's sake lets keep it at that. i'm familiar with both systems.
its a 660p. i am sometimes tired and must've been a typo. we both have the same drive, here's some box pictures: https://drive.google.com/drive/folders/12IKIKOa9oY6OXWHzoVBAqsqPS3bZbcQA?usp=sharing

A couple of maybe obvious questions. Are you running AHCI or IDE mode for your SATA drives? IDE mode would slow them down quite a lot.
Are you using GPT or MBR partitioning, as this makes a huge difference for NVMe drives.

It's ahci. i dont even have the ide option in the bios.
all drives are gpt

DISKPART> list disk

  Disk ###  Status         Size     Free     Dyn  Gpt
  --------  -------------  -------  -------  ---  ---
  Disk 0    Online          931 GB  1024 KB        *
  Disk 1    Online          119 GB      0 B        *
  Disk 2    Online          489 GB  1024 KB        *
  Disk 3    Online          953 GB  1024 KB        *

Apples to oranges, also ADATA makes good ram, unsure about SSDs, from looks of it it is old by capacity...

umm yea, i dont know about adata ram, but the ssd is fine. anyway it's the intel one i'm focused on, not the adata one.

 

[cucker tarlson]

If this drive hits 75% full the transfers go to toilet.With 200gb free you may be over that line.

 

[Valantar]

While in no way an exact or highly accurate form of troubleshooting, monitoring drive activity in Task manager could help pinpoint the bottleneck. If your Intel drive is at 100% activity when writing at 60 MB/s, then it is the bottleneck (and I would guess due to the SLC cache being minimal at that point, combined with QLC NAND's poor write speeds, along with possibly overactive garbage collection and cache flushing trying to maintain some free SLC cache despite it being very small). If the drive activity is lower than that, chances are that the bottleneck lies somewhere else or that you're looking at a confluence of several factors.

Also, as someone mentioned, is the drive connected to a CPU or chipset-based slot? AMD chipsets only provide PCIe 2.0 ports, while the CPU-connected slots are 3.0. The 660p is AFAIK a 2x interface drive, and PCIe 2.0 x2 is not very fast (still faster than SATA, but very slow for NVMe).

 

[seccentral]

While in no way an exact or highly accurate form of troubleshooting, monitoring drive activity in Task manager could help pinpoint the bottleneck. If your Intel drive is at 100% activity when writing at 60 MB/s, then it is the bottleneck (and I would guess due to the SLC cache being minimal at that point, combined with QLC NAND's poor write speeds, along with possibly overactive garbage collection and cache flushing trying to maintain some free SLC cache despite it being very small). If the drive activity is lower than that, chances are that the bottleneck lies somewhere else or that you're looking at a confluence of several factors.

Also, as someone mentioned, is the drive connected to a CPU or chipset-based slot? AMD chipsets only provide PCIe 2.0 ports, while the CPU-connected slots are 3.0. The 660p is AFAIK a 2x interface drive, and PCIe 2.0 x2 is not very fast (still faster than SATA, but very slow for NVMe).

very interesting observation.
No, it's not 100%. it's actually barely hitting 5%. hmmm
rearding slot position, i tried both with the same result.

If this drive hits 75% full the transfers go to toilet.With 200gb free you may be over that line.

i'll keep that in mind. oh well

 

[EarthDog]

my sata nvme drive

????

M.2 is the form factor. There are nvme and sata based m.2 drives with sata based drives still being limited to sata speeds....


Edit: that isnt the issue... I see now.

 

[newtekie1]

I've got the 6000P which is the same drive plus encryption, and its effing terrible for sustained writes.

Your friend might have some sort of caching going on (or a different SSD entirely, as others mention), making it appear faster than it is, the 600p/6000p drives simply cant write that fast

I wonder if Intel's chipset driver automatically activates the write-flushing policy on NVMe drives and the AMD chipset driver doesn't?

 

[Valantar]

very interesting observation.
No, it's not 100%. it's actually barely hitting 5%. hmmm
rearding slot position, i tried both with the same result.

i'll keep that in mind. oh well

Have you tried creating a temporary RAMdisk and using that as the source of your files? A quick search came up with this list of free RAMdisk software (can't vouch for the quality, of course). That would eliminate any source file bottleneck and let your SSD run free.

As for the Task Manager results, this might indicate that the QLC NAND is the bottleneck, holding even the SSD controller back and leaving it mostly idle. If the SLC cache is skipped for some reason (which is down to how the SSD firmware handles things), the controller would only be doing the work required to write as much as the flash is able to handle, while otherwise waiting and essentially twiddling its thumbs. But as I said, it might also indicate a bottleneck elsewhere (the source SSD is the most likely culprit) or a combination of two or more factors.

I wonder if Intel's chipset driver automatically activates the write-flushing policy on NVMe drives and the AMD chipset driver doesn't?

Sounds possible, but shouldn't the drive firmware be handling this regardless of platform?

 

[newtekie1]

Sounds possible, but shouldn't the drive firmware be handling this regardless of platform?

No, it is a policy you can set in device manager in Windows. It can be toggled on and off on the fly. So I wouldn't be surprised if Intel's storage software doesn't enable it by default when it detects an Intel drive to make the drive seem faster on Intel platforms.

It is also very risky, because data isn't written directly to the drive. When the policy is enabled(which disables write flushing actually) basically Windows writes data to RAM, so transfers seem fast, and the copy "finishes" quickly, but it still works in the background to write the data from RAM onto the SSD or HDD. If the PC loses power before the data is flushed from RAM, the data still in RAM is lost.

 

[EarthDog]

...and it is typically not an issue and some drives are 'supposed' to run this way IIRC.

 

[newtekie1]

...and it is typically not an issue and some drives are 'supposed' to run this way IIRC.

AFAIK, no drive is supposed to run with flushing disabled in Windows. By default the policy should be off. Write caching is on by default, but not the policy to disable cache flushing.

 

[EarthDog]

By default the policy should be off. Write caching is on by default, but not the policy to disable cache flushing.

Correct on all quoted accounts. Again though, it is typically not an issue (unless you lose power while its clearing the cache as you stated) and some drives are 'supposed' to run this way for best results. I don't recall which drive I was reviewing, likely one that doesn't have DRAM and relies on HMB???

 

[seccentral]

I got both drives.
While the old one settles at around 50MB/s the other one bursts at 900 and settles at 600. (wrote over 200GB so i suppose it's safe to say the cache was saturated at some point)
I'm starting to suspect my drive may be faulty. Although it should still be very ok since it's total writes is barely 4TB.
will update as i do more testing

after some testing i decided to backup the data and nuke the drive dd style ( dd if=/dev/zero of=/dev/mydrive ) reboot in windows and let windows reinitialize it, i'm getting similar speeds on mine as well.
...
I can't say what fixed it, but i can say i am satisfied with how it performs.

 

[darthmauldog1125]

I found this thread while trying to troubleshoot a similar issue I have. After sluggish performance on my samsung 970 pro I did a benchmark test and found my read/write times were lower than published benchmarks and samsung's stated 3500/2700 read/write sequential. the largest differences compared to published benchmark are my 4k read/write and acc.time read/write, which are 25.19mb/s read, 36.15mb/s write with acc.time of 0.108ms read, 0.111ms write; these compare to an online published benchmark i saw using the same software of 58.24mb/s read, 185.79mb/s write with acc. times 0.022ms read, 0.020ms write.

The motherboard i'm using is an ASRock x299 extreme4. I tried 2 of the 3 m.2 ports, the third not having enough room for the 970 pro, and also purchased a PCIE x16 m.2 adapter to try using one of the x16 pcie lanes. same benchmark results. Since I have an older computer next to me I decided to put the PCIE m.2 adapter into that older board, an asus maximus iv gene-z, and while the sequential speeds were choked due to the x4 pcie 2.0 lane the 4k read/writes were much faster and the acc.time was better too: 60.56mb/s read, 188.90mb/s write, with acc.times of 0.032ms read, 0.018ms write. The older machine was running windows 7 pro x64 versus windows 10 pro x64 This confuses the heck out of me. I also decided to run a test from a wintogo win10 x64 usb boot drive so that the benchmark wasn't testing a drive that was also the OS boot drive, but I got the same cruddy benchmark results for my 4k read/writes and the same slow acc.times versus online published benchmarks. So to me this points to a motherboard issue, CPU issue, or memory issue.

I also decided to run memory tests on my 4 g.skill trident Z 4x 16GB DDR4-3200 using memtest86 usb boot drive. each test took a really long time, 4-5 hours to complete. I ran twice with all four connected: 1st run contained 1 error on test 8, 1 bit; 2nd run passed. I ran again on 3 combinations, first combo passed 2 runs. 2nd combo had 1 error (test 7, 1 bit) on the 1st run but passed the 2nd, 3rd and 4th runs, the third combo passed 2 runs. I then tested using only 1 memory stick. each one passed 2 runs without errors. So possibly a memory problem but more likely a motherboard memory controller problem? unfortunately i don't have the ability to test the memory on another motherboard yet.

any thoughts on what the problem(s) is(are)?

 

[cucker tarlson]

I found this thread while trying to troubleshoot a similar issue I have. After sluggish performance on my samsung 970 pro I did a benchmark test and found my read/write times were lower than published benchmarks and samsung's stated 3500/2700 read/write sequential. the largest differences compared to published benchmark are my 4k read/write and acc.time read/write, which are 25.19mb/s read, 36.15mb/s write with acc.time of 0.108ms read, 0.111ms write; these compare to an online published benchmark i saw using the same software of 58.24mb/s read, 185.79mb/s write with acc. times 0.022ms read, 0.020ms write.

The motherboard i'm using is an ASRock x299 extreme4. I tried 2 of the 3 m.2 ports, the third not having enough room for the 970 pro, and also purchased a PCIE x16 m.2 adapter to try using one of the x16 pcie lanes. same benchmark results. Since I have an older computer next to me I decided to put the PCIE m.2 adapter into that older board, an asus maximus iv gene-z, and while the sequential speeds were choked due to the x4 pcie 2.0 lane the 4k read/writes were much faster and the acc.time was better too: 60.56mb/s read, 188.90mb/s write, with acc.times of 0.032ms read, 0.018ms write. The older machine was running windows 7 pro x64 versus windows 10 pro x64 This confuses the heck out of me. I also decided to run a test from a wintogo win10 x64 usb boot drive so that the benchmark wasn't testing a drive that was also the OS boot drive, but I got the same cruddy benchmark results for my 4k read/writes and the same slow acc.times versus online published benchmarks. So to me this points to a motherboard issue, CPU issue, or memory issue.

I also decided to run memory tests on my 4 g.skill trident Z 4x 16GB DDR4-3200 using memtest86 usb boot drive. each test took a really long time, 4-5 hours to complete. I ran twice with all four connected: 1st run contained 1 error on test 8, 1 bit; 2nd run passed. I ran again on 3 combinations, first combo passed 2 runs. 2nd combo had 1 error (test 7, 1 bit) on the 1st run but passed the 2nd, 3rd and 4th runs, the third combo passed 2 runs. I then tested using only 1 memory stick. each one passed 2 runs without errors. So possibly a memory problem but more likely a motherboard memory controller problem? unfortunately i don't have the ability to test the memory on another motherboard yet.

any thoughts on what the problem(s) is(are)?

try disabling vulnerability patches with inspectre (remember to restart the pc)
that did it in my case

check if trim is working (trimcheck)

check if intel rst is running properly

do you really mean 36mb/s 4k write or is it 136mb/s ?

 

[darthmauldog1125]

I really meant 36mb/s 4k write. super slow. I do not know how to disable vulnerability patches with inspectre, but i will try to google how to do that. How do I check if intel RST is running properly? what is strange is that I also ran the benchmark test from a new windows 10 OS build (a togo usb boot drive) and the results are basically the same as they are now. were those vulnerability patches with inspectre built into the latest windows 10 iso from microsoft?

 

[londiste]

First, the majority of this thread is about Intel's 660p that has its own issues with speed in the long run and large amounts of data at once.
Second, Spectre vulnerability patches have pretty much zero effect on a simple benchmark like this. I have a 1TB 960 Pro that gives essentially identical results to 970 Pro and its review scores and I have all the vulnerability patches installed and active. This stuff has noticeable effect only on far more stressing loads than this.

Intel RST is not relevant. This is the motherboard/chipset driver for the NVMe device. In this case, secnvme on the screenshot shows you have Samsung's driver installed which is the optimal way. You may want to check if it is the latest version - Samsung NVMe Driver from Samsung's page (https://www.samsung.com/semiconductor/minisite/ssd/download/tools/#ge_semi_anchor_stand4) but that is it.

As for the results themselves, I am not sure. Based on just benchmark results no slowdown should not be concerning or noticeable in actual usage.
The Gene-Z results are as expected - 4K matches the reviews and sequential is capped by PCI-e 2.0 speeds. That means the drive is definitely fine.

Is that a system drive? Is anything else that could use the drive closed when you are running the test?
Other drivers - especially chipset - are up to date?

On more generic troubleshooting side - is the drive overheating by any chance?

 

[darthmauldog1125]

Ok, so I tried eliminating spectre and meltdown, with restarts before each benchmark, but same results as before. I reinstalled intel RST, and also reinstalled the samsung nvme driver for the 970 pro, with restarts after installations, but both benchmarks showed the same. I downloaded the samsung magician software which showed TRIM was on.

londiste, yes, it is a system drive, which is why I made a wintogo usb boot of win 10 x64 to run windows off of a usb stick and run the benchmark on the 970 pro while it wasn't being used as a system drive. when I did that benchmark I got the same results. The temperature is normal according to samsung magician software, and currently it is in the PCIE3 slot and has a bottom fan (intake) pointed right at it; the side of the case is also off for now while I troubleshoot, so I think the 970 pro temps are fine. Chipset is up to date, I tried updating bios versions 1 by 1 until I was on the latest bios (1.8).

By chance I noticed when I run the test right after booting into windows, like a few seconds after booting, the numbers are a bit higher, not by much but they are higher than when the system has settled for a minute. So I booted into safe mode and ran the benchmark with much better speeds, still not matching the reviews and samsung stated specs but miles better than in normal boot. 4k read/write is up to 39.39mb/s and 127.75mb/s while acc. time is down to 0.04ms/0.03ms read/write. so better, not great, but definitely better. So maybe there are some windows services causing some issues, but it still doesn't explain why the benchmarks were slow while using a wintogo usb as the system drive, right?

 

[holyprof]

Looking at the two system's specs, the Intel PC has 32GB of RAM vs 16GB on the 2700X PC. Windows cache system will use RAM as a buffer (both as read and write) which speeds up things a lot - the 40GB data being copied fits alsost entirely in RAM + let's say 2GB SLC buffer of the target SSD. Also as already mentioned above, NVMe to NVMe drive will copy much faster than SATA to NVMe.

 

[darthmauldog1125]

Changed motherboards today, but same speeds. So instead of asrock x299 extreme4 I am using the evga x299 dark. So motherboard isn't the issue. I'm going to perform a clean install of windows to rule out the OS. In the attached image you'll see only sequential is checked and that's because i ran a sequential test after running all 4 tests due to read sequential showing 600-700mb/s on the first test; this test is the 970pro in m.2 slot 1 on the evga x299 dark.

 

[darthmauldog1125]

So I have performed a clean install of windows 10 x64 but same speeds. I also bought a 970 pro 512gb drive only to have the same exact issue with 4k and acc. time. I then installed the 1TB drive as a second drive to test it while my 512gb drive was being used as the OS boot, but the 1TB had the same slow 4k speeds and slow acc. times. So, I decided to install windows 7 x64 and wouldn't you know it, speeds were very fast. I'm not sure what windows 10 pro x64 OS is doing to my nvme speeds. the attached image is the benchmark on the 1tb drive while running windows 7 x64 from it.

 

[Vario]

So I have performed a clean install of windows 10 x64 but same speeds. I also bought a 970 pro 512gb drive only to have the same exact issue with 4k and acc. time. I then installed the 1TB drive as a second drive to test it while my 512gb drive was being used as the OS boot, but the 1TB had the same slow 4k speeds and slow acc. times. So, I decided to install windows 7 x64 and wouldn't you know it, speeds were very fast. I'm not sure what windows 10 pro x64 OS is doing to my nvme speeds. the attached image is the benchmark on the 1tb drive while running windows 7 x64 from it.

Looks quite fast, the speed is right where it should be for the 970 Pro, I have the same drive 970 pro 512 and win 7 and similar results. I don't know what would make it slower under 10, sorry that isn't helpful.

 

[darthmauldog1125]

Looks quite fast, the speed is right where it should be for the 970 Pro, I have the same drive 970 pro 512 and win 7 and similar results. I don't know what would make it slower under 10, sorry that isn't helpful.

I agree, in windows 7 the benchmark is where it should be. something in windows 10 is causing things to slow down. in windows 10 safe mode boot the speeds are faster than in normal mode boot. i tried disabling antivirus, firewalls, and also tried disabling/enabling running services one by one but nothing would increase the benchmark speeds.

 

[Athlonite]

I agree, in windows 7 the benchmark is where it should be. something in windows 10 is causing things to slow down. in windows 10 safe mode boot the speeds are faster than in normal mode boot. i tried disabling antivirus, firewalls, and also tried disabling/enabling running services one by one but nothing would increase the benchmark speeds.

Just out of curiosity what build version of Windows 10 X64 pro are you using

 

[darthmauldog1125]

I used the latest windows media creation tool from Microsoft’s website to install windows 10 x64 from usb. It is Microsoft Windows 10 Pro version 10.0.18362 build 18362

 

[Athlonite]

so your saying it's build 1903 (19H1) hmmmm weird I've not seen any major complaints for storage speed NVMe or otherwise under this build also weird that it works as it should under Win7

could be a firmware issue have you looked at whether or not there's a new firmware update for your 970 pro on Samsung's website