|
If you’re a data hoarder you can find some old Cliffdale 8TB drives for sub $400 https://serverpartdeals.com/product...e97a1e3c6&_ss=r Cliffdale is poo poo tier enterprise drives but probably fine for your Plex server
|
#
¿
Oct 2, 2023 00:01
|
|
|
|
| # ¿ May 13, 2024 20:19 |
|
|
Got back from Samsung Memory Day I hope you fuckers are excited for QLC in phones
|
|
#
¿
Oct 21, 2023 05:33
|
|
|
why spec for 5 years endurance when people replace their phones after 3.5 years or less and new phones come out every year
|
|
#
¿
Oct 21, 2023 18:34
|
|
|
hug your nearest solidigm friend because they are doing mass layoffs right now
|
|
#
¿
Oct 25, 2023 00:27
|
|
|
Not sure but they force retired some department head and force merged two disparate groups so they are probably doing a huge cull
|
|
#
¿
Oct 25, 2023 00:34
|
|
|
Nothing takes the string out of having a meeting scheduled for the next day where you know you will get fired like calling that meeting “Departure Support”
|
|
#
¿
Oct 25, 2023 00:39
|
|
|
I don't know NAND that well so I'm poorly (mis)remembering things told to me or from materials I read, but there is complexity in adding layers. Basically the vertical channel they have to etch/drill down all the way to the bottom of the layers is difficult, you can't have it twist or taper too much and I'm assuming if one of the layers is bad then the whole die is lost. String stacking was kind of a hack to get around this by stacking and connecting dies and getting the subsequent layer count. Adding layers is important because that's how you increase your capacity. Micron's 6500 ION is TLC but matches Solidigm's QLC drive in size simply because of the layer advantage. In terms of QLC, I also know very little but Solidigm from what I can tell positions themselves as the QLC Champions (no one is going bigger than them or pushing QLC in the data center as much). But they've abandoned (or been forced to abandoned by Korean management) floating gate and off-hand remarks I've heard made it sound like FG couldn't really scale beyond ~200ish layers (Solidigm QLC I believe is at 192). QLC isn't incompatible on charge trap but I think people were curious as to what the hell they are going to do since their QLC endurance characteristics may take a step back if they have to move to CT (and the FG roadmap is effectively dead). A 200+ layer QLC drive is interesting from a potential density point of view but if the endurance is even worse than a typical QLC drive then probably less interesting. Typically, the high layer count/leading edge stuff when it first comes out is sold as client/consumer products as the endurance is lower and you want to flex that cost advantage of the higher layer count. Data center/enterprise typically lags in adoption as they try to work the kinks out and get the endurance up to the expected level. Micron kind of short cutted this with the ION though as they realized they could still get the endurance up to the level of QLC enterprise drives while still matching in size (and blowing away in performance) What is interesting is a Chinese NAND company matching or exceeding the layer count of the big dogs. In spite of tech bans too. I suspect many NAND company ceos wake up in cold sweats at the idea of cheap Chinese nand flooding the market WhyteRyce fucked around with this message at 21:06 on Oct 25, 2023 |
|
#
¿
Oct 25, 2023 17:31
|
|
|
Wow Solidigm is letting everyone who gets fired to keep their laptops what a swell gesture
|
|
#
¿
Oct 25, 2023 20:19
|
|
|
priznat posted:How do they do that, just remote wipe? Remote wipe and go buy your own Windows license it sounds like Or comedy option put Ubuntu on it
|
|
#
¿
Oct 25, 2023 20:26
|
|
|
Manager: but don’t people we end up hiring to backfill these losses need a laptop?! Solidigm: hahahaha
|
|
#
¿
Oct 25, 2023 20:29
|
|
|
For fun you can do this yourself with spdk, a quarch, and manually managing the queues yourself MY PRODUCTS DIDNT HAVE THIS ISSUE!!! WhyteRyce fucked around with this message at 05:56 on Nov 22, 2023 |
|
#
¿
Nov 22, 2023 05:41
|
|
|
priznat posted:I wonder if the failing drives all use the same controller https://twitter.com/xenadu02/status/1496290369184874497
|
|
#
¿
Nov 22, 2023 05:57
|
|
|
it doesn't sound like he's harvesting and verifying the queues which is absolutely needed to say the drive is losing data due to power loss. but it wouldn't surprise me if there is a miss by some controllers since power loss verification is hard and I suspect vendors focus more on the "kill power and make sure drive comes back alive" and less on the "verify the accepted transactions are flushed". the former is easy to blast away at with fio and a quarch, the later is more difficult to setup. not sure if applies to client drives that have no cap, but technically if the controller can guarantee what's in the cache will be written to media during power loss (all enterprise drives), then it is considered non-volatile and flush command is a no-op WhyteRyce fucked around with this message at 10:30 on Nov 22, 2023 |
|
#
¿
Nov 22, 2023 06:18
|
|
|
Even if Intel client SSDs still existed they’d be SMI based stuff anyway
|
|
#
¿
Nov 28, 2023 17:47
|
|
|
if you’re buying third party turn key solutions you probably can save engineering head count by sticking to the reference firmware except in “value add” areas (i.e. random nvme and security features, nothing performance related). To a bean counter it really doesn’t make sense to invest in an in house full firmware team if you’re already buying a complete solution. Also funny things happen when you decide to do that anyway and the firmware team brazenly ignores reference design and recommendations Client is a poo poo razor thin even by nand product standards market and with the exception of a few standout cases I don’t think their client strategy was anything other than a numbers game to keep their fab utilization high or at least soak up the nand that wasn’t good enough quality for the enterprise market they actually cared about WhyteRyce fucked around with this message at 08:44 on Nov 29, 2023 |
|
#
¿
Nov 29, 2023 08:38
|
|
|
priznat posted:It’s true, stuff like the 750 and 900p were too good to justify as you could charge the same for way shittier products. AND they would have much better margins as you barely change the firmware off a mid tier supplier’s reference design. The 900p was fun because it was mostly the exact same as the enterprise 4800x with some firmware features disabled (namely manageability and logging stuff). But that was probably just the byproduct of the "JUST BUILD SOMETHING AND FLOOD THE MARKET" mindset of early Optane because the follow on product had no client/gamer/enthusiast variant.
|
|
#
¿
Nov 29, 2023 20:22
|
|
|
I’m not too familiar with channel prices but it looks like you can get 2 8TB TLC drives for the price of a 16TB QLC drive if you can manage that I was assuming the p5316 would crater now that p5336 is out but those prices are still high. If you can hilariously manage an E1.L in your setup you can get a 16TB p5336 for just under a grand With the micron and solidigm nvme drives you can also force them to a lower power state if you want to save power at the cost of performance. I think both let you go down from 25W to around 10W WhyteRyce fucked around with this message at 22:04 on Dec 16, 2023 |
|
#
¿
Dec 16, 2023 22:00
|
|
|
E1.L is the same physical connector as E1.S so if you don’t mind ugly rear end macgyver rigs you can probably find some cabling adapter solutions. Word of warning, those e1.s to u.2 adapters you can find online only work if there is no heatsink on the drive Another word of warning, e1.l drives are taller so it probably won’t fit in an actual proper e1.s enclosure At some point I hope p5316s flood the market. They really aren’t bad for Plex drives, even with the hilariously large indirection size WhyteRyce fucked around with this message at 02:51 on Dec 17, 2023 |
|
#
¿
Dec 17, 2023 02:29
|
|
|
Since this is an academic question: Depends on what you mean by small files. Most drives that I'm aware of have a 4KB indirection unit, so each location in the mapping table corresponds to a 4KB sized location. If you're writing 4KB or higher, things line up nicely. If for some reason you're doing 512B file writes, that's a lot of read-modify-write going on which is obviously bad for WAF. Probably rare that this happens in most cases. But on a more extreme end of that, the Solidigm P5316 QLC drive has a hilarious 64KB indirection unit. You can see the potential problems you get with that. It's honestly an ok drive for what it is but please stick to large file writes or you get bad performance and worse endurance. Aligning to your IU size is key, but enterprise high capacity drives are probably the only drives that are the odd duck here so don't panic over your client drive. How you write also matters. If you look at the high capacity low endurance Solidigm QLC and Micron 6500 drives, you see that they start breaking out the drive writes per day rating between random and sequential. These values can differ greatly (like 3x in the 6500 case) so they do it to highlight the best way to write to the drive and sell you on the concept they aren't as fragile as you'd think WhyteRyce fucked around with this message at 17:27 on Dec 22, 2023 |
|
#
¿
Dec 22, 2023 17:05
|
|
|
If you’re going with a weirdo setup then lean the gently caress into it and get a 4 port mini sas pcie switch and roll with u.2 drives Something like this works but you can maybe find it cheaper elsewhere https://www.amazon.com/dp/B089SQ84CH?ref=ppx_pop_mob_ap_share
|
|
#
¿
Jan 1, 2024 04:50
|
|
|
Also they absolutely do make enterprise m.2s but they are 22x110 and sadly seem to be a dying breed No clue if they are as plentiful on the lightly used market as u.2s
|
|
#
¿
Jan 1, 2024 04:57
|
|
|
Samsung PM9A3s seem to be availble on Ebay in the open box/lightly used variety Primary concern with Samsung enterprise drives is I don't think the FW binaries are publicly posted online (for enterprise lines you're supposed to contact your channel/sales support) and the Magician software won't update it either (although maybe the enterprise toolkit will do the same). Not sure if other vendors have the same policy. Micron appears to put theirs online but finding lightly used/open box 7450s seems harder to do lol someone dumped the pm9a3 firmware files and changelog here. for the curious, feel free to read the fw release notes and see a detailed description of what types of bugs you can hit on SSDs https://c.gmx.net/@324567863383819227/ooFXSOMeTcu9nk4ziFVa8w https://forums.servethehome.com/index.php?threads/pm9a3-firmware-errormod-related-to-fw-version.36822/ WhyteRyce fucked around with this message at 05:40 on Jan 1, 2024 |
|
#
¿
Jan 1, 2024 05:18
|
|
|
Hasturtium posted:I don’t have a promiscuous amount to throw away on this, and this combo seems about my speed. Does this seem tenable? pretty much any pcie -> u.2 adapter card I've used works no problem, including the cheapo ~$10 ones on amazon behold this beauty for half the price of the other switch chip I linked! https://www.sfpcables.com/pcie-to-sas-expansion-card-pcie-x8-to-4x-sas-sff-8643-connectors WhyteRyce fucked around with this message at 05:44 on Jan 1, 2024 |
|
#
¿
Jan 1, 2024 05:39
|
|
|
BobHoward posted:I haven't even gotten to reading the bugs yet but am floored by the fact that they send out release notes as excel spreadsheets A different company sends them out as .ppts 
|
|
#
¿
Jan 8, 2024 06:29
|
|
|
I very much empathize with FW engineers having to work on Opal stuff. With the NVMe spec, if you implement something "wrong" you can just hem and haw about specification grey areas and why what you're dong is actually valid, push out a FW fix if that doesn't work either, ask for a waiver if that doesn't, and, worst case, tell a customer they are SOL if all else fails. People working on Opal are probably the same class of engineers but get none of the leeway. Also they are under the same project schedule time crunch and resource constraints as everyone else, maybe even more since no one probably wants to work on it
|
|
#
¿
Jan 8, 2024 06:49
|
|
|
Anime Schoolgirl posted:these things are now showing up on Amazon for $430 and now i'm weighing whether I should continue with my plan to get the 960gb 905P for a hundred dollars less just to own a piece of history or get the P4510 for the staggering amount of space (because I don't have the budget for both) This is only $350 for 8TB from Micron drive https://www.crucial.com/ssd/7400_pro/mtfdkcb7t6tdz-2az18abyyr
|
|
#
¿
Jan 30, 2024 16:19
|
|
|
The nand price bounces are a real thing although I’ve heard people point out that vendor reaction to the cycles exacerbates the issue
|
|
#
¿
Feb 3, 2024 19:58
|
|
|
Solidigm sells their large capacity qlc drives in u.2 not just e1.l so you don’t have to go that route unless it’s a cost thing As far as the adapters, you need to be careful with what you get. I bought a cheapo gen z to u.2 adapter since e1.l won’t fit in e1.s slots on my existing systems (e1.l isn’t just longer). Turns out a lot of the early gen z adapters were for bare PCB products. The adapters have some screw mounts that prevent you from using a drive that has a heat sink (i.e. all of them). So you’re only options are to remove the heat sink (probably bad idea) or remove the screws and mounts and hope that the drive doesn’t fall out of the adapter slot Example of said adapter: https://a.co/d/7ZH5Fr5 Clicking the link now though shows this guy as a other buying option which might actually work https://a.co/d/cVxZlxU Given e1.l length, if you’re using a regular pc instead of an actual server or something without u.2 slots, the above adapter with a u.2 cable and a pcie to u.2 adapter may give you needed flexibility in drive placement WhyteRyce fucked around with this message at 00:16 on Feb 6, 2024 |
|
#
¿
Feb 5, 2024 23:44
|
|
|
Seems to be available other places Oh wait that’s 4TB
|
|
#
¿
Feb 6, 2024 21:44
|
|
|
If your SSD is in need of a FW patch then things are wrong to start
|
|
#
¿
Mar 17, 2024 00:48
|
|
|
Having done both I’m feeling very triggered right now
|
|
#
¿
Mar 17, 2024 04:28
|
|
|
A certain SSD vendor will never admit you found a bug and will instead insist it’s the specs fault for not allowing fault tolerance and special messages to mitigate said bug. It’s exhausting
|
|
#
¿
Mar 17, 2024 04:32
|
|
|
there are a lot of reasons why Optane died. the tl;dr though is basically cost/capcity. you were basically paying more for a drive that was 1/10th the capacity. optane gets a lot of poo poo for failing but I think that's a tad unfair. it's the same poo poo computational storage, K-V SSDs, storage accelerators, ZNS, and probably FDP will struggle with. yes, tremendous benefit can be squeezed out under optimal, well written code. but at the end of the day the customers all just seem to want something that's cheap as poo poo, works like everything else, and doesn't require a rewrite of whatever poo poo software stack they are having to deploy. having to explain "literally none of your old code will run on this ZNS drive unless maybe you switch over to this other experimental file system also your performance may be poo poo" to a customer probably instantly sours them. WhyteRyce fucked around with this message at 17:25 on Mar 17, 2024 |
|
#
¿
Mar 17, 2024 17:22
|
|
|
Optane DIMMs were regular dram backed by Optane so it was like the memory hierarchy within a memory hierarchy if you used the right mode It’s not just persistent, it’s huge amounts of memory. I’ve seen enough demos and presentations to know that there is a potentially valuable performance improvement to be had that isn’t just no need to reload from disk after reboot, but yeah rewriting anything is a hard sell even if it’s for potentially good reasons. ZNS seems pretty dead in favor of FDP because of the hard requirement of a complete rewrite even though the former gives you a real cost advantage. And I think many companies weren’t keen on jumping all in to a single sourced tech from a historically finicky vendor There still is interest around the concept. It’s why CXL is forging ahead. WhyteRyce fucked around with this message at 15:47 on Mar 18, 2024 |
|
#
¿
Mar 18, 2024 15:32
|
|
|
I don’t even know FDP is that much easier (the spec is kind of junky tbh) but it lets you use it as a traditional SSD in parallel to or as backup to your application teams figuring out just what the hell they are supposed to be doing so maybe that’s enough
|
|
#
¿
Mar 18, 2024 16:56
|
|
|
I used to think nothing of stuff like dual sockets being an issue for anyone because I’m a brain dead HW guy but then I talked to application teams that absolutely hate numa and lol CXL is numa on steroids get in losers we are going over an interconnect
|
|
#
¿
Mar 18, 2024 18:20
|
|
|
nvme drives are not required to support the sanitize operation, or any the specific sanitize operations. Always need to check the identify controller output for support which is probably not easily doable in windows really surprised someone would not crypto erase though you should still be able to kick off secure erase and specify crypto vs. user data erase via nvme format though WhyteRyce fucked around with this message at 00:47 on Mar 20, 2024 |
|
#
¿
Mar 20, 2024 00:39
|
|
|
|
| # ¿ May 13, 2024 20:19 |
|
|
Self encryption does not mean Opal. In fact don't lots SSDs leave their internal encryption on because there is almost no overhead but you get the benefit of scrambling data and also the benefit of fast wipe? But yeah I guess not every ASIC may implement it which explains my question but I was always told the you get free (and necessary) benefits out of it so that throws me off but my head is stuck in enterprise land I've got a stack of non-Opal drives that support crypto erase WhyteRyce fucked around with this message at 02:10 on Mar 20, 2024 |
|
#
¿
Mar 20, 2024 01:34
|
|