question Defective SSD

[BORODA]

Hello,

I recently bought an ADATA NVME SSD ADATA Swordfish 2TB.

Installed it into a fresh new build based on GB-Z690 UD DDR4 as a secondary drive for games. I already have another NVME SSD installed for system and it works just fine.

Swordfish is plugged into second NVME port in the mobo (chipset).

I haven't used it much, but recently I noticed that sometimes the read speeds drop to abysmal levels, like 300 kb/sec and task manager shows 100% disk load, and performance monitor has a disk queue length of 5.

 

When I installed the drive, I didn't do the "full" format, just a quick one. Install, create partition, quick format. So I decided to move the files away from it and do a full format of the drive.

At that moment I noticed that the format speed is fine up to 15%. The performance monitor shows 1400 mb/s write speed before the formatting reaches 15% and 20 mb/s after it reaches the 15%, and it goes like this afterwards. I checked the temperature of the drive with CDI - it sits around 52-55 C, so it's not overheating. Also I tried the SSD toolbox and checked for firmware update. The toolbox says that there's a new firmware available, but when I download the utility they make you run, it says the drive is up to date.

Is the drive dead?

Should I try to return it?

Or it could be something else?

 

Thanks.

[Levent]

14 minutes ago, BORODA said:

1400 mb/s write speed before the formatting reaches 15%

Which is the expected behavior. Drive is not dead, it could be the combination of type of files you are transferring, running the drive on chipset and being a bad performing drive overall.

 

What does Crystal Disk Mark show?

[BORODA]

12 minutes ago, Levent said:

Which is the expected behavior. Drive is not dead, it could be the combination of type of files you are transferring, running the drive on chipset and being a bad performing drive overall.

 

What does Crystal Disk Mark show?

Why would it be an expected behavior for a linear write which formatting is? I know that random write could be slow, but this is a all-linear write of "zeroes" to disk. And the random drops in reads that the drive experiences sometimes which renders the drive completely unusable. And by unusable I really mean unusable, i.e. if it happens during the game loading, the load will not happen, I waited for 10 minutes yesterday for a AoE4 game to load before giving up. The task manager showed 100% drive load and 300 kb/sec read speed.

 

Crystal Disk Mark show around 1800 mb/s read and 1600 mb/s write on SEQ1M Q8T1.

I don't have other results at hand but they are all within the spec of the drive.

[Vishera]

24 minutes ago, BORODA said:

Why would it be an expected behavior for a linear write which formatting is?

Once the DRAM cache is full the drive will be significantly slower.

[BORODA]

1 minute ago, Vishera said:

Once the DRAM cache is full the drive will be significantly slower.

This drive does not have a DRAM cache, and 20 MB/s? This is slower than my USB 2.0 stick.

And, this does not make sense for reading.

[Vishera]

1 minute ago, BORODA said:

This drive does not have a DRAM cache, and 20 MB/s? This is slower than my USB 2.0 stick.

And, this does not make sense for reading.

That's just how cheap DRAM-less QLC SSDs perform.

[BORODA]

2 minutes ago, Vishera said:

That's just how cheap DRAM-less QLC SSDs perform.

20 mb linear write and 300 kb random read is "expected"?

[jaslion]

3 minutes ago, BORODA said:

20 mb linear write and 300 kb random read is "expected"?

For a cacheless qlc drive yeah not that unexpected. I've seen my 860qvo do worse.

[Vishera]

7 minutes ago, BORODA said:

20 mb linear write and 300 kb random read is "expected"?

Yes,It is expected.

 

Just return it and buy something from SAMSUNG's EVO series or KIOXIA's EXCERIA PLUS series.

[BORODA]

16 minutes ago, Vishera said:

From this model - Yes,It is expected.

The controller on those is pretty bad.

 

Just return it and buy from SAMSUNG's EVO series or KIOXIA EXCERIA PLUS series.

Oh. Damn.

I will not be able to return it. I'm not in country where laws work and unless it's broken or defective if the box is open I'm SOL and stuck with it

 

  

18 minutes ago, jaslion said:

For a cacheless qlc drive yeah not that unexpected. I've seen my 860qvo do worse.

This is not even a QLC drive, it's stated as TLC. I specifically filtered out QLC drives when I was looking for a budget option to run games from.

[jaslion]

24 minutes ago, BORODA said:

Oh. Damn.

I will not be able to return it. I'm not in country where laws work and unless it's broken or defective if the box is open I'm SOL and stuck with it

 

  

This is not even a QLC drive, it's stated as TLC. I specifically filtered out QLC drives when I was looking for a budget option to run games from.

It's tlc but it seems to be very slow tlc. From reviews the behavior you find here is mostly replicated.

[mariushm]

Your SSD has a portion of its memory set in pseudo-SLC mode, where a single bit is written in each memory cell, instead of 3 bits for the TLC memory or 4 for the QLC memory chips. By using that memory in SLC mode, the writes are faster and that memory has higher endurance (let's say 10-15k erases vs 1-2k erases for TLC or <1k for QLC). 

As you write stuff on your SSD, the controller writes the data into that fast SLC cache area, and when it has time, it gradually empties the cache in background, moving the data to TLC or QLC memory.

So your 15% of 2 TB is around 75 GB ... seems reasonable amount of SLC memory for a 2 TB drive. 

 

When you're formatting (which is pointless for SSDs), the format application writes each "sector" filling it with information and then the format application tries to read that sector and check that the written information was written correctly ... so you're basically filling that write cache at very high speed and then the controller is forced to write directly into the TLC or QLC memory at the slower speed of the memory (usually should be 300-500 MB/s and for QLC speeds in the 20-50 MB/s are possible)

 

Also, SSDs don't work like mechanical drives ... data is not arranged linearly like it is on mechanical drives, and this is one of the reasons defragmenting is pointless on SSDs, because the SSD controller fragments the file content on purpose

When the format application tells the ssd to fill the first megabyte with zeros or whatever, the SSD controller has to look up in what memory chip, in what block of memory is that first megabyte translated to.. then when the format application asks for 2nd megabyte ... the ssd controller has to ask again.. and that 2nd megabyte could be in another chip, or another layer of the chip.

 

[BORODA]

36 minutes ago, jaslion said:

It's tlc but it seems to be very slow tlc. From reviews the behavior you find here is mostly replicated.

It's odd to see ADATA Falcon in Tier-C (Mid-End) in this case.

Because I'm looking at the reviews now and the falcon seems to be even worse than Swordfish...

[BORODA]

22 minutes ago, mariushm said:

Your SSD has a portion of its memory set in pseudo-SLC mode, where a single bit is written in each memory cell, instead of 3 bits for the TLC memory or 4 for the QLC memory chips. By using that memory in SLC mode, the writes are faster and that memory has higher endurance (let's say 10-15k erases vs 1-2k erases for TLC or <1k for QLC). 

As you write stuff on your SSD, the controller writes the data into that fast SLC cache area, and when it has time, it gradually empties the cache in background, moving the data to TLC or QLC memory.

So your 15% of 2 TB is around 75 GB ... seems reasonable amount of SLC memory for a 2 TB drive. 

 

When you're formatting (which is pointless for SSDs), the format application writes each "sector" filling it with information and then the format application tries to read that sector and check that the written information was written correctly ... so you're basically filling that write cache at very high speed and then the controller is forced to write directly into the TLC or QLC memory at the slower speed of the memory (usually should be 300-500 MB/s and for QLC speeds in the 20-50 MB/s are possible)

 

Also, SSDs don't work like mechanical drives ... data is not arranged linearly like it is on mechanical drives, and this is one of the reasons defragmenting is pointless on SSDs, because the SSD controller fragments the file content on purpose

When the format application tells the ssd to fill the first megabyte with zeros or whatever, the SSD controller has to look up in what memory chip, in what block of memory is that first megabyte translated to.. then when the format application asks for 2nd megabyte ... the ssd controller has to ask again.. and that 2nd megabyte could be in another chip, or another layer of the chip.

 

I know the thing about cache and I wouldn't actually noticed the slow write speed if it wasn't for atrocious reads. As I mentioned above, I couldn't load an AoE4 match for about 10 minutes before I gave up and rebooted because the drive was at 100% with 300 kb/s reads.

I didn't expect the random reads to bottleneck the drive that hard. I expected it to be "slow" compared to faster SSDs (frankly, my system one is ~3x faster), but I didn't expect it to be slower than mechanical drive during read.

 

I didn't know that linear write out of cache would lead to 20 mb/s writes, which is terrible, but again, I'm fine with it as long as reads are okay'ish.

[BORODA]

I got back home and if the rtl_nvme_flash_id.exe utility is something to believe, this drive actually has QLC NAND. Seems like ADATA is lying...

v0.14a
OS: 10.0 build 22000 
Drive   : 2(NVME)
Scsi    : 2
Driver  : W10
Model   : ADATA SWORDFISH                         
Fw      : VC0S0302
HMB     : 32768 - 65536 KB (Disabled, 64 M)
Size    : 1907729 MB [2000.4 GB]
LBA Size: 512
Fw Str  : [REALTEK_RL6577 ] []
Bank00: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank01: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank02: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank03: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank04: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank05: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank06: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank07: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank08: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank09: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank10: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank11: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank12: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank13: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank14: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank15: 0xad,0x89,0x2c,0x53,0x30,0xa0,0x0,0x0 - Hynix 3dv5-96L QLC 16k 1024Gb/CE 1024Gb/die
Bank32: 0xd1,0x9f,0x68,0xd6,0x1a,0x90,0x31,0xa7 - 
Bank33: 0xb9,0x24,0x10,0xa3,0x4d,0x5f,0x91,0x7d - 
Bank34: 0xe9,0xe6,0x7d,0x82,0x31,0xdd,0x1b,0x75 - 
Bank35: 0x80,0xf6,0xba,0x0,0xa1,0x87,0x70,0xe7 - 
Bank36: 0xc7,0x48,0x70,0x1e,0x86,0xaf,0xaa,0xfd - 
Bank37: 0x92,0xed,0xa5,0x88,0x83,0x99,0xb5,0xc3 - ESMT(PowerChip) MLC
Bank38: 0x13,0xd3,0xad,0xe8,0x3f,0x86,0xe5,0x25 - 
Bank39: 0x9c,0x2d,0x2b,0x3,0xe4,0xe9,0xb1,0xc5 - 
Bank40: 0x2c,0xd9,0xeb,0x64,0xd6,0xdc,0x58,0x90 - Micron TLC 64Gb/CE 8Gb/die
Bank41: 0xc4,0xfd,0x5c,0xeb,0x9f,0x83,0xa1,0xe1 - 
Bank42: 0x89,0xa6,0x18,0x92,0xaf,0x7f,0x79,0x89 - Intel TLC
Bank43: 0xd0,0x61,0xdb,0x20,0x6c,0x58,0x80,0x78 - 
Bank44: 0x9e,0x79,0xea,0x7e,0xa,0x56,0x3c,0x34 - 
Bank45: 0xc1,0x40,0x27,0x2d,0x73,0x60,0xab,0x89 - 
Bank46: 0x1f,0x1b,0x8,0x49,0xf,0xa2,0xff,0xac - 
Bank47: 0x23,0x9e,0x31,0xba,0x74,0xf2,0x78,0xa7 - 
Bank48: 0x28,0xb,0xac,0xa5,0x10,0xfc,0x9b,0x42 - 
Bank49: 0x95,0x22,0x2f,0x19,0x2b,0x84,0xa6,0x83 - 
Bank50: 0x94,0x89,0xd3,0x22,0xc2,0x7d,0xad,0x5f - 
Bank51: 0x59,0x86,0x53,0x31,0x34,0x2f,0x56,0x2a - 
Bank52: 0xbf,0x91,0xce,0x28,0x1f,0xb8,0xf8,0x6e - 
Bank53: 0x51,0xab,0x99,0x31,0x2d,0x71,0x1f,0x12 - Qimonda TLC
Bank54: 0x75,0xec,0x2f,0x77,0x98,0x40,0x72,0xf5 - 
Bank55: 0xfb,0x47,0x63,0x79,0x3f,0x54,0x54,0x54 - 
Bank56: 0x7c,0xf4,0xc1,0x6a,0x1c,0x6b,0x13,0xa8 - 
Bank57: 0xaf,0x51,0xf6,0xb3,0x13,0xf,0xe,0xd2 - 
Bank58: 0xa,0xe,0x8e,0x57,0xfe,0x26,0xf1,0x15 - 
Bank59: 0x43,0x1c,0xda,0xf4,0xb9,0x61,0x2a,0xb5 - 
Bank60: 0x86,0x4,0x55,0x95,0x5,0x3f,0x97,0xb3 - 
Bank61: 0xfc,0xbf,0xce,0xb7,0x22,0xe8,0xa9,0xaf - 

 

[BORODA]

This is what it looks like. Steam launch takes around 2 minutes. This is slower than any mechanical drive I ever owned I believe.

 

Spoiler

 

Is this really considered normal?

There's nothing else loading from the drive, only steam starting up (there's nothing on that drive except for steam actually).