Please Stop Buying the Wrong SSD

[Shawn Frey]

What makes an SSD good or bad? Did you know there's an important hardware difference between top-performing and low-cost storage devices? You might not, because many SSD makers don't talk about DRAM.

[vonKordke]

If you don't know the difference and you can't tell without watching the video, why bother? Ignorance is bliss, especially when you spent less money

[CyberneticTitan]

Where's the benchmarks? So many performance claims made in this video without data.

[steamrick]

WD / Sandisk also quietly put revisions of the Ultra 3D / Blue 3D SATA SSDs on the market that completely removed the DRAM cache.

Probably when the SA510 came to market and they just quietly substituted the newer, cheaper-to-produce internals.

 

Do you also remember when WD started putting SMR inside its red braneded NAS drives?

 

Yeah.

 

PS: If you mention HMB, please also remember to point out that it only works with NVMe drives, never through SATA or USB.

[dogwitch]

Surprise this was not brought up in video.  for ref .

[Tan3l6]

7 minutes ago, dogwitch said:

Surprise this was not brought up in video.  for ref .

Yeah, Adata SX8200 Pro was so much "revisioned" it somewhat started to feel like a real scam.

 

SSD tiering lists have gone extinct mean time.

[starsmine]

Not a single benchmark to be found. Basing the recommendations on what was true 6 years ago. Not a single data point to be found anywhere in this video to back up the claims.

HMB is more then enough for any modern system including the OS drive, it does not work in the PS5 sure, but this was about PCs. 

[dogwitch]

11 minutes ago, Tan3l6 said:

Yeah, Adata SX8200 Pro was so much "revisioned" it somewhat started to feel like a real scam.

 

SSD tiering lists have gone extinct mean time.

that and mobo manf lying with pci lanes on a mobo (linus talk about the issue years ago)

 

[Kododile]

2 hours ago, vonKordke said:

If you don't know the difference and you can't tell without watching the video, why bother? Ignorance is bliss, especially when you spent less money

True. I always just get whatever is cheapest from samsung, crucial or like another big brand i recognize. 

[ThankGodItsFriday]

How ironic that this video should be posted today. I'm in the process of replacing a TeamGroup SATA M.2 drive that just suddenly died, and I suspect that it's because it has little to no on-board protections that prevent premature hardware failure. I'll be sure to watch this soon.

[cmndr]

The 400GB SSD I just bought for a boot drive doesn't have a DRAM cache at all =(. 
What do you guys think about an Intel Optane p5801x. 

[starsmine]

21 minutes ago, cmndr said:

The 400GB SSD I just bought for a boot drive doesn't have a DRAM cache at all =(. 

If its an nvme drive on a modern system, usign HMB, its a complete and total non issue.

21 minutes ago, cmndr said:

What do you guys think about an Intel Optane p5801x. 

why do you have one of those laying around? yes those are good drives, its optane not flash, nothing in the video even applies to you. 

[Lurking]

5 hours ago, steamrick said:

PS: If you mention HMB, please also remember to point out that it only works with NVMe drives, never through SATA or USB.

This. If you later use that SSD in an USB enclosure, you will regret skimping on DRAM. 

 

Even as m.2, accessing system RAM means "reaching" through the CPU into system AM and all the way back. 

 

The only manufacturers that don't change parts/specs after the "glorious" reviews are out are Samsung/SK Hynix. Linus is right, even WD don't ensure what actually is in the SSD. But he is wrong about relying on Techpowerup, since those data could be outdated (unless they do frequent re-confirmation...). But he is right, Samsung actually puts it on the spec sheet, which makes it legally binding. WD just promises the SSD is black, and they keep that promise  If a manufacturer isn't willing to put specs on paper, you know what is up. 

[cmndr]

3 hours ago, starsmine said:

If its an nvme drive on a modern system, usign HMB, its a complete and total non issue.

why do you have one of those laying around? yes those are good drives, its optane not flash, nothing in the video even applies to you. 

I was being a smartass. 
It's arguable that a p5801x is functionally the fastest drive on the market for OS use. No DRAM cache because the 3D Xpoint chips are fast enough to not need it. 

 

I originally bought it as an L2ARC device for a NAS (Minisforum MS-01) and then I discovered that it physically won't fit... so I plopped in a 280GB optane 900p and the p5801x is going into my desktop (someday... still waiting on an adapter to arrive from China). 

[bluehawk]

Maybe LTT can build a database with their labs site so when we go to buy one we will know. 

[Capt Fiero]

Anybody watch Project Farm, or Torque Test Chanel.  Every week, just simple no BS testing. No bias, just raw data.     I am even a paying member for TTCs database of tools.   Its really helpfull when Dewalt has 6 difrent 1/2 drive impacts for sale all that look similar, but one part number off means you could be getting 500lbs of torque vs 1000lbs of torque in your impact gun.     We need something like that for PC components.   I'd love to know price to performance metrics on all this stuff.   I'm like many people I try to buy a brand name part but with a reasonable price, and just hoping that the one I am grabbing A. works good, and B. lasts a least a few power cylces.   My last M.2 Drive, was working great one day, and dropped dead the next, my entire OS drive just died, no warning.   I even tried dropping it into an M.2 to USB C enclosure.  I was able to get a small bit of data off of it, oddly enough via my Samsung Note20 cell phone when windows could not see it at all.    I only really needed one file, outlook.pst  everything else was replaceable.    Yes I am aware of the CPU and GPU benchmark sites and I routinely check cards I find on marketplace against them, and sadly found out just how Bad my spare Linux boxes CPU really is.  (Intel 750) 

[mariushm]

The video is just bad and full of misinformation. 

 

Sometimes a controller that support RAM cache is discontinued and has to be replaced with a more modern controller that supports higher pci-e speeds and newer flash communication protocols.  It's not a race to the bottom all the time. 

 

Let's compare the WD models ,  SN750 (2019) , SN750SE (2021) 

 

SN750 : https://www.techpowerup.com/ssd-specs/western-digital-sn750-1-tb.d460

SN750SE : https://www.techpowerup.com/ssd-specs/western-digital-sn750-se-1-tb.d455

 

SN750 came with a 8 channel controller with 3 cores, 28nm, which peaked at 5.8w power consumption, and does only pci-e 3.0 and nvme 1.3

 

The new one has a cheaper 4 channel controller that peaks at 4.4w and support nvme 1.4 and pci-e 4.0

 

Also, the old controller only supported a maximum of 14 GB of static SLC memory (portion of TLC flash that's always kept in SLC mode until the drive is nearly full) while the more modern Phison controller supports up to 70 GB of SLC flash, which is converted to TLC as the drive fills up.  So with the newer model, you can write BIG files for longer time, before speed drops. 

 

On the old drive, copy a 50 GB file and after 14 GB, speed drops to 1.6 GB/s, on the newer drive you'd have to transfer 70 GB until the speed drops to 1 GB/s 

 

So this change cuts costs in microcontroller and reduces the amount of heat produced on the SSD, and you only get a reduced performance if you copy very large files (larger than 70 GB). 

 

The reduction in channels is not a big deal, because they moved from 64 layer TLC chips to 96 layer chips and higher capacity chips, so you get higher read/write speeds from each memory chip - only on sustained very long writes the reduction to 4 channels will be noticed in the difference in speed (1.6 GB/s vs 1 GB/s after the 70 GB cache is filled)

 

The old flash chips were :

Technology:     64-layer, Speed:     533 MT/s Toggle: 2.0 19 nm Die Size:     132 mm² (3.9 Gbit/mm²) Die Read Speed:     400 MB/s Die Write Speed:     46 MB/s
Endurance: (up to)     3000 P/E Cycles (30000 in SLC Mode) 

 

The new flash chips are : 

Technology:     96-layer Speed:     800 MT/s ONFI:     4.0 Toggle:     3.0  19 nm Die Size:     86 mm² (6.0 Gbit/mm²)  Die Read Speed:     551 MB/s Die Write Speed:     57 MB/s Endurance: (up to)     3000 P/E Cycles (30000 in SLC Mode) 

 

So same endurance, faster read and write speeds due to the 96 layers, which means faster writes in the SLC cache portion, as long as the 70 GB write cache is not filled the drive IS better. 

 

The lack of DRAM only hurts if you have really high random read/write, because the latency of reading data from ram through the pci-e bus to read the lookup table from HBM memory is higher than reading from local dram chip, but the memory chip is expensive and also an extra memory controller inside the controller chip would increase the costs and die size making it more expensive.  

 

 

[Timme]

13 hours ago, starsmine said:

Not a single benchmark to be found. Basing the recommendations on what was true 6 years ago. Not a single data point to be found anywhere in this video to back up the claims.

HMB is more then enough for any modern system including the OS drive, it does not work in the PS5 sure, but this was about PCs. 

Dude, seriously, pick up any online review of the DRAM drive. This video is not it. Or read some articles about the DRAM vs *less. Don't you have any sense of inquisitiveness?

[starsmine]

1 hour ago, Timme said:

Dude, seriously, pick up any online review of the DRAM drive. This video is not it. Or read some articles about the DRAM vs *less. Don't you have any sense of inquisitiveness?

The video is what LMG published. 

Also 

Quote

HMB or DRAM-less SSDs have been refined so much that most normal users would never be able to experience any practical performance difference.

But again, that DRAM vs *less article has fuck all in terms of their own data to back that up. 

[Timme]

1 minute ago, starsmine said:

The video is what LMG published. 

Also 

But again, that DRAM vs *less article has fuck all in terms of their own data to back that up. 

What would you consider a valid backing up data? Design dock for the hardware/firmware interactions? You won't see that because trade secret. Only a general description of functions by manufacturers, or press professionals in the field. Tests, again.

But why does the notion of having a dedicated fast cache storage chip on the drive not seem like an improvement to you, anyway? HMB is only 64 MB, which is nowhere near enough to store a map of an SSD with data on it. Also, DRAM cache can store data for quick access (write or pre-caching). With all that, one other great benefit of DRAM is that it reduces wear on the NAND itself when you deal with an OS drive.

[leadeater]

12 hours ago, Lurking said:

This. If you later use that SSD in an USB enclosure, you will regret skimping on DRAM. 

 

Even as m.2, accessing system RAM means "reaching" through the CPU into system AM and all the way back. 

 

The only manufacturers that don't change parts/specs after the "glorious" reviews are out are Samsung/SK Hynix. Linus is right, even WD don't ensure what actually is in the SSD. But he is wrong about relying on Techpowerup, since those data could be outdated (unless they do frequent re-confirmation...). But he is right, Samsung actually puts it on the spec sheet, which makes it legally binding. WD just promises the SSD is black, and they keep that promise  If a manufacturer isn't willing to put specs on paper, you know what is up. 

With the very few exceptions of I don't care and I expect this to fail I only buy Samsung SSDs and will continue to only buy Samsung SSDs. I might be leaning SUPER hard on that trust and knowing how Samsung SSDs are made and that they don't make different revisions of controllers and NAND (if they do it meets the original spec) and different parts always get named as a new different part i.e. VNAND generation packages. I or Samsung has yet to do or be wrong about this.

 

I've probably overspent, I'm satisfied with that.

[DeerDK]

I'm starting to think getting my 4tb Samsung 990 Evo plus wasn't such a great buy. 

But I got it before the prices went crazy, so at least I've got plenty of storage for not so much cash 

[dogwitch]

20 hours ago, dogwitch said:

Surprise this was not brought up in video.  for ref .

 

11 hours ago, Capt Fiero said:

Anybody watch Project Farm, or Torque Test Chanel.  Every week, just simple no BS testing. No bias, just raw data.     I am even a paying member for TTCs database of tools.   Its really helpfull when Dewalt has 6 difrent 1/2 drive impacts for sale all that look similar, but one part number off means you could be getting 500lbs of torque vs 1000lbs of torque in your impact gun.     We need something like that for PC components.   I'd love to know price to performance metrics on all this stuff.   I'm like many people I try to buy a brand name part but with a reasonable price, and just hoping that the one I am grabbing A. works good, and B. lasts a least a few power cylces.   My last M.2 Drive, was working great one day, and dropped dead the next, my entire OS drive just died, no warning.   I even tried dropping it into an M.2 to USB C enclosure.  I was able to get a small bit of data off of it, oddly enough via my Samsung Note20 cell phone when windows could not see it at all.    I only really needed one file, outlook.pst  everything else was replaceable.    Yes I am aware of the CPU and GPU benchmark sites and I routinely check cards I find on marketplace against them, and sadly found out just how Bad my spare Linux boxes CPU really is.  (Intel 750) 

ref video above is why

[starsmine]

3 hours ago, Timme said:

What would you consider a valid backing up data? Design dock for the hardware/firmware interactions? You won't see that because trade secret. Only a general description of functions by manufacturers, or press professionals in the field. Tests, again.

But why does the notion of having a dedicated fast cache storage chip on the drive not seem like an improvement to you, anyway? HMB is only 64 MB, which is nowhere near enough to store a map of an SSD with data on it. Also, DRAM cache can store data for quick access (write or pre-caching). With all that, one other great benefit of DRAM is that it reduces wear on the NAND itself when you deal with an OS drive.

SSDs for the last decade dynamically switch to SLC mode for writes. There is no wear concerns for SSDs in any consumer sense, even with massive page files. you can write the whole drive over 3 times a day and it will still last 5 years before eating into the overprovisioned flash. 

HMB is enough to store any warm map of the SSD. Cold maps literally do not mater. 

What I'm saying isn't to avoid DRAM cached drives, what I am saying is the video is unnecessarily disparaging to dram-less drives with zero data to back it up. 

The god damn xda press professional article you link has the same god damn problem. ZERO data to back up their claims and then they link to a previous article with again, zero data to back up their claims. They keep perpetuating something that was true A DECADE AGO as still true today. And as of 2024 in windows, HMB is no longer just 64MB.

I am not here saying DRAM drives are not better, what I am saying is that its a marginal gain that 98% of users, even hardcore power users. will ever notice. If linus wants to argue that DRAM drives are better though in a published video. have some evidence to back it up is my major complaint (as well as my complaint with all those opinion articles with zero tests backing it up)

Yes, the drives are worse when in a USB enclosure. but honestly the only times I do that is for recovery of a system. I am not expecting high performance in general form any type of USB enclosure, I think the stronger argument would be thunderbolt. 

There are tools out there to help them build benchmarks for this like intel's MLC for their lab https://www.intel.com/content/www/us/en/developer/articles/tool/intelr-memory-latency-checker.html
 

[Timme]

1 minute ago, starsmine said:

SSDs for the last decade dynamically switch to SLC mode for writes. There is no wear concerns for SSDs in any consumer sense,

Are you serious right now? How tf does SLC caching prevent wear?
And, again, what "actual data" do you need? So far, there are benchmarks, and the internet is full of reputable sources posting articles on the matter. How about this: it is you who is wrong and doesn't have any data to back that up.