4 ms·

We only have 4 manufacturers of SSD drives who also making their own chips: Intel, Micron, Samsung, and SK Hynix. Here’s the table: https://en.wikipedia.org/wik

We only have 4 manufacturers of SSD drives who also making their own chips: Intel, Micron, Samsung, and SK Hynix. Here’s the table: https://en.wikipedia.org/wiki/List_of_solid-state_drive_manufacturers https://en.wikipedia.org/wiki/List_of_solid-state_drive_manu... I tend to buy Samsung and Intel, both working well in my devices.

kogepathic 5y ago
I thought Western Digital had vertically integrated their SSDs after the purchase of SanDisk? According to that Wikipedia list, Western Digital does manufacture their own flash, through a joint venture with Kioxia.
- andruby 5y ago
  Also, Western Digital is looking at completely buying that Kioxia https://www.reuters.com/technology/western-digital-advanced-talks-merge-with-japans-kioxia-holdings-wsj-2021-08-25/ https://www.reuters.com/technology/western-digital-advanced-...
justinclift 5y ago
> I tend to buy Samsung and Intel, both working well in my devices. It's a generalisation, but be careful of using new Samsung consumer SATA drives (860, 870) if you're using Linux. The Linux kernel has workarounds in place for the older ones (840, 850 series). But the workarounds for the newer models are only starting to be looked at this week: https://bugzilla.kernel.org/show_bug.cgi?id=203475 https://bugzilla.kernel.org/show_bug.cgi?id=203475 Without those workarounds (eg disabling NCQ trim), there are data loss problems.
- Const-me 5y ago
  Thanks for the warning. Fortunately, I’m unaffected because running Windows. I’m running Linux on top in WSL1 and VMWare. Also running Linux on ARM devices which have neither SATA nor PCIx. When I need fast disks in them to compile my C++ code, I simply use a network-mounted drive to build, gigabit LAN is good.
- wtallis 5y ago
  > disabling NCQ trim Fixed support for queued trim was part of the marketing claims for the 860 series. Whatever's going on with the 860 and newer drives requires deeper investigation, because there's plenty of reason to believe that simply applying the old workaround is not accurately targeting the real problem.
  - justinclift 5y ago
    The old workaround was to turn off support for NCQ (queued) trim, instead just using standard trim. It doesn't seem to have any real downsides. :)
    - wtallis 5y ago
      > It doesn't seem to have any real downsides That depends greatly on what your software's policy for issuing Trim commands is. Partly because of the dearth of drives properly supporting queued trim, it is common for Linux distro to only run trim commands via cron job, rather than issuing them in real time during a workload.
- alias_neo 5y ago
  Thanks, I had no idea about this. I've been using Samsung drives in all of my machines and servers and I run Linux exclusively. With a mixture of 850, 860 and 870's I must be at least partially affected.
- aidenn0 5y ago
  Are 860s new? They are what's in my system and while I don't remember when I upgraded it was definitely pre-COVID... [edit] smart says they have 631 days of power-on time, so maybe 2019? [edit] I seem to remember having issues a while back but I haven't seen any since I upgraded my firmware to RVT04B6Q
  - justinclift 5y ago
    k, I probably should have said "newer" (than 850 model) rather than "new". The 870 model though, is pretty new. :)
scns 5y ago
Sadly my goto Crucial started doing it too, but they paired it with a huge price drop. Was the P2 NVME IIRC. ~70€ for 1TB.
scns 5y ago
Just read that Samsung changed the BOM on 970 evo plus. Switched to the controller of the 980 pro. Test reults are mixed. 4K QD1 is up ~50->82MB/s, serial writes after exhausting SLC cache (115GB) are nearly halved 1500->800MB/s.
- wtallis 5y ago
  That's not what happens when you switch to a newer, faster controller. Those performance changes are what happens when you switch to newer NAND flash that's manufactured with twice the per-die capacity, so you have half as many dies for a given drive capacity. (Assuming the newer NAND isn't split into twice as many planes per die to compensate.) Edit: Looking into this further, TechPowerUp, Computerbase and others have mistaken K90UGY8J5B for K9DUGY8J5B. It's an easy mistake to make, and I've done exactly this before. But it completely explains the performance difference. The D that changed to a 0 signifies a switch from 16 dies per package down to 8 dies per package. The digits signifying the capacity of the package have stayed the same. The "B" at the end signifying the generation has also stayed the same, but Samsung didn't introduce 512Gbit TLC dies until a generation after they introduced 256Gbit TLC dies, so on the smaller dies the "B" means 92L and on the larger dies the "B" means 128L.
  - kasabali 5y ago
    > That's not what happens when you switch to a newer, faster controller. Wasn't also caching strategy changed with the newer controller? Larger dynamic slc, slower folding speed?
    - wtallis 5y ago
      I haven't done the math, but my gut feeling is that the relatively modest increase in SLC cache size could not on its own lead to that big a drop in post-cache sustained write speed except possibly where QLC is concerned, and especially not when mitigated by a generational increase in per-die performance. Samsung's drives are still operating with SLC caches that are smaller than the theoretical limit, which means that even after the cache is full, the drive should be able to accept some amount of writes that bypass the cache and have near native TLC performance; flushing the cache can be deferred a bit longer until the drive is actually starting to run out of free blocks, as opposed to typical QLC drives with maximally-sized SLC caches that run out of free blocks at the same point where the cache runs out.
randerson 5y ago
And Intel just sold their SSD business to SK Hynix, so soon there will only be 3!
fomine3 5y ago
This issue happened on WD (acquired SanDisk). So the "just buy NAND manufacturer's SSD" strategy don't work. Also you miss Kioxia!

kogepathic 5y ago

I thought Western Digital had vertically integrated their SSDs after the purchase of SanDisk? According to that Wikipedia list, Western Digital does manufacture their own flash, through a joint venture with Kioxia.

andruby 5y ago
Also, Western Digital is looking at completely buying that Kioxia https://www.reuters.com/technology/western-digital-advanced-talks-merge-with-japans-kioxia-holdings-wsj-2021-08-25/ https://www.reuters.com/technology/western-digital-advanced-...

justinclift 5y ago

> I tend to buy Samsung and Intel, both working well in my devices. It's a generalisation, but be careful of using new Samsung consumer SATA drives (860, 870) if you're using Linux. The Linux kernel has workarounds in place for the older ones (840, 850 series). But the workarounds for the newer models are only starting to be looked at this week: https://bugzilla.kernel.org/show_bug.cgi?id=203475 https://bugzilla.kernel.org/show_bug.cgi?id=203475 Without those workarounds (eg disabling NCQ trim), there are data loss problems.

Const-me 5y ago
Thanks for the warning. Fortunately, I’m unaffected because running Windows. I’m running Linux on top in WSL1 and VMWare. Also running Linux on ARM devices which have neither SATA nor PCIx. When I need fast disks in them to compile my C++ code, I simply use a network-mounted drive to build, gigabit LAN is good.
wtallis 5y ago
> disabling NCQ trim Fixed support for queued trim was part of the marketing claims for the 860 series. Whatever's going on with the 860 and newer drives requires deeper investigation, because there's plenty of reason to believe that simply applying the old workaround is not accurately targeting the real problem.
- justinclift 5y ago
  The old workaround was to turn off support for NCQ (queued) trim, instead just using standard trim. It doesn't seem to have any real downsides. :)
  - wtallis 5y ago
    > It doesn't seem to have any real downsides That depends greatly on what your software's policy for issuing Trim commands is. Partly because of the dearth of drives properly supporting queued trim, it is common for Linux distro to only run trim commands via cron job, rather than issuing them in real time during a workload.
alias_neo 5y ago
Thanks, I had no idea about this. I've been using Samsung drives in all of my machines and servers and I run Linux exclusively. With a mixture of 850, 860 and 870's I must be at least partially affected.
aidenn0 5y ago
Are 860s new? They are what's in my system and while I don't remember when I upgraded it was definitely pre-COVID... [edit] smart says they have 631 days of power-on time, so maybe 2019? [edit] I seem to remember having issues a while back but I haven't seen any since I upgraded my firmware to RVT04B6Q
- justinclift 5y ago
  k, I probably should have said "newer" (than 850 model) rather than "new". The 870 model though, is pretty new. :)

scns 5y ago

Sadly my goto Crucial started doing it too, but they paired it with a huge price drop. Was the P2 NVME IIRC. ~70€ for 1TB.

scns 5y ago

Just read that Samsung changed the BOM on 970 evo plus. Switched to the controller of the 980 pro. Test reults are mixed. 4K QD1 is up ~50->82MB/s, serial writes after exhausting SLC cache (115GB) are nearly halved 1500->800MB/s.

wtallis 5y ago
That's not what happens when you switch to a newer, faster controller. Those performance changes are what happens when you switch to newer NAND flash that's manufactured with twice the per-die capacity, so you have half as many dies for a given drive capacity. (Assuming the newer NAND isn't split into twice as many planes per die to compensate.) Edit: Looking into this further, TechPowerUp, Computerbase and others have mistaken K90UGY8J5B for K9DUGY8J5B. It's an easy mistake to make, and I've done exactly this before. But it completely explains the performance difference. The D that changed to a 0 signifies a switch from 16 dies per package down to 8 dies per package. The digits signifying the capacity of the package have stayed the same. The "B" at the end signifying the generation has also stayed the same, but Samsung didn't introduce 512Gbit TLC dies until a generation after they introduced 256Gbit TLC dies, so on the smaller dies the "B" means 92L and on the larger dies the "B" means 128L.
- kasabali 5y ago
  > That's not what happens when you switch to a newer, faster controller. Wasn't also caching strategy changed with the newer controller? Larger dynamic slc, slower folding speed?
  - wtallis 5y ago
    I haven't done the math, but my gut feeling is that the relatively modest increase in SLC cache size could not on its own lead to that big a drop in post-cache sustained write speed except possibly where QLC is concerned, and especially not when mitigated by a generational increase in per-die performance. Samsung's drives are still operating with SLC caches that are smaller than the theoretical limit, which means that even after the cache is full, the drive should be able to accept some amount of writes that bypass the cache and have near native TLC performance; flushing the cache can be deferred a bit longer until the drive is actually starting to run out of free blocks, as opposed to typical QLC drives with maximally-sized SLC caches that run out of free blocks at the same point where the cache runs out.

randerson 5y ago

And Intel just sold their SSD business to SK Hynix, so soon there will only be 3!

fomine3 5y ago

This issue happened on WD (acquired SanDisk). So the "just buy NAND manufacturer's SSD" strategy don't work. Also you miss Kioxia!