Et tu, Samsung? Electronics giant accused of quietly switching SSD components Samsung has altered the parts used to make its 970 EVO Plus 1TB SSD card, leading a version manufactured in June 2021 to perform differently than an older formulation from April 2021. In a video posted to the channel YouTube channel 潮 玩 客 ("Trendy Player"), Chinese video blogger Jian Ge recently compared two versions of the …
Those big corporations don't have competition, so they can effectively form cartels and do whatever.
You could technically build your own SSD if only the chips were available on the open market.
Maybe governments should force companies to not only enable component manufacturers to sell parts to general public, but also force them to release documentation.
Imagine you could get PCB files (gerbers, pick and place configuration) from Github and just submit an order to one of fabs and get your own drive. More ambitious could could just assemble at home (you can get all tools under £1000 and you could probably get return of your investment by selling spare drives you made).
Imagine if Apple was forced to release M1 chip and its documentation to the open market.
> For many processors you can get all documentation ...
I tried to get a full datasheet for an Atom processor off Intel. I had to register, which was a load of faff. My request for the data was turned down, which could have been for any number of reasons. They weren't specific.
I don't usually get that kind of runaround. I just download the PDF off the website. Mostly, I don't even have to register on the site.
Because selling chips and documenting them is unprofitable and would ruin a company. That's why outfits like AMD, Intel, Silicon Motion, Micron, Texas Instruments, Renesas, Cypress, Fujitsu, Arduino, Adafruit, 4D, Omron, Phoenix, Apacer, Winbond, Xeltek, TDK, Kioxia, Digilent, Everspin, Fluke, Bridgetek and so forth don't exist.
I work at a company that builds it's own servos and control systems. If you know what you are doing, you don't want someone else's gerber files and such. What you want is proper chip data sheets and a stable supply chain. We presently have suppliers quoting over 70 week lead times on some FPGA and SOC chips. SSD's are pretty simple circuit-wise, which is why if you look at one there's not much to it. The firmware which handles wear leveling and self-encryption and the like is where DYI SSD's would probably have an issue trying to do it right..
“ Imagine if Apple was forced to release M1 chip and its documentation to the open market.”
Imagine if the reference designs that apple uses are available from the vendor and anyone could buy them?
Imagine if the very people who make those designs could help you customise to your requirements and there was some other business that would fab for you too?
Now what’s the barrier for the ordinary person on the street doing what apple do from the same people apple use?
If you’re reliant on a non-published spec of a product and the manufacturer changes the ingredients such that they are still meeting their published specs then you’re taking a risk, even between batches of the same model. I think Sammy’s actions are almost totally fine, but.
It’s a royal pain for folk trying to maintain homogeneous infrastructure or replace like-for-like. So, put an additional number at the end of the model number, how hard can that be (SAP be damned).
EVO960-blah-b
EVO960-blah-c
Still hits all the searches, the marketing material is all still good, avoids this sort of ‘are they trying to dupe their customers’ vibe.
My Dell i7 laptop came with a just about adequate sized SSD, which started to fail just out of warranty. I have no idea of the speed or spec beyond actual capacity- which goes for most users outside, I'd assume corporate IT teams and the like.
What I do know is that the Dell PC had been pretty sluggish compared to what I'd expected- until I slapped in the new SSD (Samsung as it goes).
Suddenly it was giving me all the performance I'd expected from an i7 with plenty of RAM.
The DELL came with an SSD described by capacity, just as a HDD would have. Clearly though some beancounter somewhere had decided to cut a few corners.
There is more to a product's capabilities than what its manufacturer advertises. If you read reviews and see benchmarks, you expect it to hit ALL those benchmarks, not the limited set the manufacturer advertises. If you buy one and it performs at a certain level for your application / usage model, you expect to be able to buy another one and hit the exact same level of performance.
If vendor X is selling a smartphone and advertise it as having x megapixels and a screen of a certain size and so forth, you would be unhappy if they changed the lens for the camera to one that was lower quality or changed the display to one that had a big issue with reflectivity in sunlight, even if it otherwise met all the manufacturer specs on their web site.
This "it hits all the specs" sounds like a Samsung apologist trying to claim what they did was somehow not as bad as the other SSD vendors.
My main worry would be more regarding the driver. I remember back in the day when Netgear used to change the controller chip in their network kit without changing the product name. Some versions worked under BSD, some did not, so it was a pain to ensure you were purchasing the correct one.
As it stands, without revision, etc,looks more like fraud that anything else,bait and switch.
Tbf, they should be different models, the difference between a 970 and a 960 is mostly controller and nand.. and some of these companies are changing controller, nand and firmware. So a completely different product
Samsung makes its own flash chip, that claim (chip shortage) is ridiculous. Especially when upgrading to the newer VNAND! They're short of old chips, therefore they make new chips to fill the gap and a new design and set up the new production line and new testing?? Cos of a temporary chip shortage, in the other Samsung factory....the one that has to make the new chips.
It seems to be a lot faster for most use cases, not sure why're your're bitching. Anything under 160GB in a single burst and its a lot faster and bigger writes you're unlikely to be keep it doing a solid burst for. Even if you can 1TB burst, its 11 mins vs 19 minutes, which is still a lunchtime backup.
I haven't done the weeding in the garden yet, its the chip shortage! I know the bike needs it tyres pumped up, and when the chip shortage ends I'll get right on it. Bins? Chip shortage! Fix the wash...CHIP SHORTAGE!
Samsung makes its own flash chip, that claim (chip shortage) is ridiculous. Especially when upgrading to the newer VNAND! They're short of old chips, therefore they make new chips to fill the gap and a new design and set up the new production line and new testing?? Cos of a temporary chip shortage, in the other Samsung factory....the one that has to make the new chips.
it's not just NAND, it's also the controller chips they are having a shortage of. They changed the controller from the Phoenix controller to the Elpis controller that they use on the newer 980 series. Samsung makes a significant percentage of it's SSD controllers in the Austin Fab in Texas. You know, the one that was shutdown due to the power outages in January, and took more than 2 weeks to get back online. This severly hampered Samsungs SSD controller production on top of the general supply issues caused by COVID.
Therefore yes, they build their own controllers (and NAND), but one of those specific controller fabs got shutdown for a substantial (in chip production terms) period of time, from this article (emphasis mine) :
That would put the full shutdown time at about three weeks. That would correspond to a worldwide 3-5 percent drop in wafer starts in February. That’s not much more than a blip in normal times, but at a time when every facility is under pressure, this is unwelcome news.
Still, IMO, rather than using components used in other SSDs in the 970, they should have just stopped - or reduced - 970 production and increased 980 (since they obviously have 980 controllers and NAND to spare to put in the '970') until they had fixed their supply issues.
If a fab loses power unexpectedly, most "in process" work is lost and the wafers are scrap. Probably they had enough advanced notice they were able to complete all the in-progress steps (some of which take maybe a day or so) to avoid this, but we don't really know that for certain.
Wafers can take several months from the time they enter the first step to the time they exit the last step, so if you take a hit at the exact wrong time you could potentially lose 10-20% of your entire year's output even if the power was only off for a day.
You're half right, pretty sure Samsung is still the fab for those chips, as you yourself pointed out. Just because they license a design doesn't mean they're impacted by a shortage in that sense. What is probably impacting them, though, is the shortage of raw silicon stock they need as input to their fabs.
I wonder how many people are actually affected by this particular change? I don't deal with production kit/servers/big systems so based on;
"starts at about 2,500 MB/s and maintains this until writing about 115 GB of data, after which performance falls to 800 MB/s."
How many and what sorts of applications would be writing a 115GB stream of data frequently enough to be an issue? I'd imagine there are many people who will see the faster initial write speed, and never hit the "wall" where it slows down. I'm assuming here that the 115GB "wall" is a cache, and if you stop writing before hitting it that the cache then empties, hopefully before the next big write. I can imagine that for many uses, the uplift in initial write speed and the larger cache before speeds drop will actually be a benefit. The problem to me, seems to be the lack of transparency in the changes in specs and the lack of a new model#.
Thinking of backups, where do they hide the UPS to make sure all the data that was streamed into the cache (and claimed to be written) has made it through to the (possibly drastically slower) SSD chips before a power fail.
Which leads me to wonder whether the read of long requests is similarly diddled. Given the nature of the actual chips, I suspect "maybe, but not near as much".
But there is no such thing as a backup if you can't restore.
The cache is two tier for 970 EVO Plus, there's 512MB-2GB of DRAM (depending on drive capacity) followed by 13-42GB of SLC (again based on capacity) followed by the stated capacity in 3D TLC in the models reviewers got their hands on. Assuming it's still the same in the new models since it wasn't mentioned as a change.
It doesn’t have to make it through. If these guys have say an 80GB cache, completely full when power fails, then when power is restored, the drive has an 80GB cache full with data and will start immediately to write it to the bigger, slower components. That’s no problem whatsoever. Or do you think they have 80GB of RAM?
> what sorts of applications
People doing video editing can easily be dealing with TB-size media files. 8k, 40 minute HDR video would be several hundred GB's (depending on compression levels, even bigger if lighly compressed, even RAW).
Many big (and not so big, especially if they are enthusiastic about video making technology) youtube channels are already recording in 8k.
And if you are editing many streams at once, that's a fuckton of data.
"that's a fuckton of data"
I can only dream of tossing something like that at YouTube.
I still make my videos in 720p because out here in the middle of nowhere, my upstream is about 65kbytes/sec, so a 1.2G twenty minute video takes a long time to upload (in the ballpark of four hours).
Various rural development plans keep promising fibre, in "about three years". They've been saying this since around 2012. I think they keep underestimating how complex (and expensive) it will be to hook up a million farms and houses where many of them are not close to each other.
Thankfully, it seems as if the government has enacted a non-discrimination policy, else the telco would probably just wire up the towns and call it job done. So... Nice fast fibre in about three years. And I kind of suspect in about three years I'll tell you the same thing!
In the case of that kind of data you’re more likely to be using a storage array rather than a single drive, in which case how much difference does it really make? If you want to smash around 8k video you know there’s a very high system cost to accompany that.
In that application the first 115 GB take 46 seconds to copy, and the second and any further 115 GB take 144 seconds. Around 100 seconds wait until the full speed is restored. So you can copy 115 GB in 46 seconds once every 144 seconds.
Some people will hurt, some will never notice with their usage. Samsung could sell drives without those limits under a new name for more money - I assume they cost more to build.
Hadn’t noticed that the original drive had lower cache speed and a tiny cache. My calculation is that the new drive writes 115 GB substantially faster (46 vs 74 seconds) and writes 180 GB in exactly the same time. So at least 95% of customers get a better drive for their purposes.
> what sorts of applications would be writing a 115GB stream of data frequently enough to be an issue
People who record large chunks of RF spectrum for minutes/hours/days/weeks at a time around the world (I'm not saying who would need to do such a thing, but there are very few customers who can legally buy the products due to ITAR/EAR restrictions). Some SDR (Software Defined Radio) hardware could sample 16GHz RF bandwidth at 16-bit samples, that would be the In-phase and Quadrature components, so you would be talking about a constant data stream of 64GB/second!
e.g. https://www.pervices.com/
And Per Vices also offer to replace the stock ADS54J60 (dual channel 16-bit 1GSPS) in their cyan product with ADC32RF45 (dual channel 3GSPS @ 12-bit only). So in theory you might need to record 144GB/second if you were using all 16 IQ channels at once.
That's an interesting description of what some people do. But these drives max out at 2500MB/s write speed so I'd guess the people you are talking about are going to be using something a bit more beefy like full on enterprise grade kit writing into a large array and will be testing and benchmarking and carefully checking the specs of the kit they are using or planning to use. Would someone writing 64GB/s continuously for days or weeks really be using 1TB NVMe drives? I just can't imagine that amount of data or begin to conceive of the costs involved, it's just so far outside of my experience.
People who play video games? Seeing as 115GB is below the size of most AAA title data sections these days you'd likely feel it when doing something as simple as shuffling games around on Steam. Although to be fair you'd have to go well past 115GB for the average speed to drop, and the initial speed is now higher.
I think the key issue is that they send these things out for reviews by consumer review outlets (Tom's, GN, whoever) that they know are going to measure and publish all this stuff. So sending one batch out for review then immediately switching to a different batch that actual consumers will receive is shady as hell regardless of the impact of the specific change. It's a practice that needs to be stomped down.
We musos have has put up with this sort of thing since forever. Guitar pedals are low man on the totem pole when it comes to components, so chances are that lovely sounding pedal you had but foolishly sold 20 or more years ago costs more than your house, because the chips inside are made from unobtanium. Newer versions don't sound the same (I'll leave out the subjective better/worse), so what you buy as a replacement isn't what you wanted.
A good example is the old Electro Harmonix Electric Mistress, which used SAD1024 BBDs as part if its characteristic sound. When these things were no longer available, EHX switched to other BBDs which, to many - including me - sounded nowhere near as good.
Even well established companies such as Boss swap components from time to time; the venerable DS-1 has been through at least 4 revisions and many others have gone the same way. Even now, one of their early digital delays goes for silly money because of the legendary long chip it contains.
The HDD shall rise again !
(icon because data are our present $eity)
...now that I think about it, it's not too late to switch to papyrous...
- long-term detention of data - check
- constant performance characteristics over long periods of time - check
- cheap long-term storage - check
- readily available - check
- simplicity of read/write mechanisms - check
- fault-tolerance - check
If it's not the same thing internally, why does it have the same reference? If the design has had to be changed for $Reasons, can't they be a little more honest looking and bump the 970 to be the 971, and so on? Or, at the very least, the 970-2 or 970b or something.
They should have advertised the new model as enhanced version with faster write speeds. They could even gain by being be honest about the slow-down after 115 GB and tell shoppers to look for the old version if the 115 GB speed-bump would be problematic. Everybody loves transparent performance data and choice! Almost everyone would buy the new version and be happy about the better performance.
I think you GREATLY overestimate 'data' and 'choice' from the general public. G. public don't give a f... about specific data (never mind the question of what it means in real-life operation), but they care greatly about BIGGER! FASTER! SMOOTHER OPERATION!
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
