I don't have a good understanding of NAND chips. For reference, how does this one compare to other NAND chips?
400MB/s peak per chip doesn't sound that fast, because joining 8 of them together would only result in 3200MB/s. (Sequential or Random not specified) This seems average when a Samsung 980 Pro 1TB will hit up to 7000MB/s sequential read/write.
Now, if that 400MB/s per chip was random read/write, then that would be very impressive.
"Now, if that 400MB/s per chip was random read/write, then that would be very impressive."
it's never been clear to me why sequential and random don't, by the arithmetic, work out to be (nearly) identical. we know, don't we?, that NAND/SSD data isn't stored contiguously as it is on disks, but spread all over the NAND, and importantly, re-spread during operation vis wear-leveling. so how does sequential access measure better?
Might be the sustained data rate, and 3.2GB/sec would be amazing for that. Most of the sustained read/write fall off a cliff when hybrid SLC cache is exhausted. Some of my NVME drives still can't keep up with the sustained throughput of my 970 Pro SSD on large file transfers for instance. Probably the worst I've seen is a 1TB NVME I use for backup data that falls to about 100MB/sec sustained once the cache is exhausted.
>it's never been clear to me why sequential and random don't, by the arithmetic, work out to be (nearly) identical.
Because of latency. There's a ~100us delay between when you start a read and when data is available. For 4KB reads at QD1 (meaning each waits for the one before to complete), that is 10,000 per second, or 40MB/s. Going faster means lower latency and NAND is optimized for density/bandwidth but not latency.
In short, the controller can't do as many small block pulls and set them up exactly right. It's not quite the same as seek time but it is really about the controller not the medium. If you look at high end enterprise controllers and ssd's the rates are much closer.
Sorry to put a finer point on it....there is no such thing as a random read speed really. It's latency on the random then you start reading, then you do a find and read again. So random is usually not measure in read speed but rather the latency to find the new file/parts of a file and then it starts to read as fast as it can. Depending on the size of the file you run into efficiencies of file systems, operating systems, and controllers but generally it's two completely different things.
According to specification, ONFI uses 8 or 16bit data bus. This multiplies with 3.2GT and 4 channel can give you up to 12.8GBps bandwidth. Theoretically you can have a low end dramless 4 channel controller hitting that speed. Or more than pcie 5.0x4 can provide if you use an 8 channel controller
Great a die shrunk TLC NAND chip. I'm sure the lack of detail about P/E endurance is not at all an indicator of further reductions in useful lifespan at all. Nope, not even a little bit. Instead please pay attention to performance specs.
These are just models and extrapolations. Now that we've had consumer SSDs for 15 years it's time someone published concrete real life data.
The Tech Report RIP did an SSD endurance test back then, but conducting retention tests is much difficult.
(If you go looking for the retention articles, brace yourself. The site was purchased a few years ago and turned into a random crypto-gambling-clickbait cesspool.)
A 2020 summary for results between 3d-TLC and QLC flash ("The errors and the optimal read voltages of 3D TLC and QLC flash-memory chips are explored.") might be a starting point ( https://www.microarch.org/micro53/papers/738300a48... ) although it seems pretty theoretical results considering looking for consumer friendly review or summary on data retention parameters and storage treatment recommendations. I do agree, public data on data retention seems rare considering market share of SSDs nowadays.
back2future & sheh Thanks for trying to look into this matter. I did try and research it a few years ago with about the same success you two have had. The 2015 AT article was from one manufacturer (Intel) many many years ago and I think it was for planar flash, and before stacking, and before QLC came out, and before we changed the cell type to charge trap flash.... so not really that useful.
The 2018 QLC comment section and 2021 FerriSSDs articles are just some off hand comments and an advertisement of a product intended for the auto market. The JEDEC standard I think is JESD219A. It costs as much as a 1TB PCIe 4.0 flash drive to obtain, just so I know about how long my SSDs will retain their data -- and I probably have to sign an agreement not to tell anyone else about data-retention. When will the tech world learn to give us some hard data?! What happened to science being important?! Is it only for their exclusive club?!
The Infineon paper you cite is talking about floating gate flash which, AFAIK, is not used in modern consumer flash products. They don't test anything in their other than storing the drive at 85C, which how long it retains data is based on how many times it was P/E (Programmed and Ereased).
Regarding your IEEE article, that's the most interesting of the bunch. I'll have to read that later...
EDIT: Regarding the Infineon paper, (AFAIK) it's 85C they're testing at, 55C real world operating conditions. I.E. cold storage isn't taken into consideration -- and as you might recall from the Intel post you cited from AT, 55C is the absolute best case scenario. Even Intel, back then, got ~8 years of data retention in that bast case scenario. See section 3.1, third paragraph from the bottom.
2015 article is real information with look-up tables from JEDEC standard for hardware that time (if there's smth. like that for 202x QLC Nand flash comparing development of retention and hardware qualities was possible?), 2018 shows steady interest on retention, but only on commenters sections (being one example on that maybe biased view) and FerriSSDs being one example for technology dev.s to remedy short data retention specs (another was reheating flash cells having high P/E cycle numbers for curing defects and somewhat extremely extending P/E cycles, IIRC from 1-10k (~MLC or TLC flash classes) to <1-100M, but only theoretical considerations made on that approach) 55°C is highest temperature for active device on JEDEC (Intel) table, enabling highest ~400 weeks data retention (on SLC/MLC types?) with steady(?)/average 25°C storage temperatures. Me, no flash gate differences, nor production node sizes considered with researching some overview information. Infineon and IEEC abstract correlate on information towards each other (~1k years endurance for 1time or low W/E cycles, but seems nonsense expecting linearity for a 1k years retention within diagram figure6, i'ld absolutely agree on that point) for Infineon with probably 2d SLC or MLC flash cells (10k-100k P/E endurance mentioned). What's data retention for HDD's, current USB flash devices, sd-cards or M.2&NVME Nand storage devices (, DVD or Bluray)?
and a short view on an abstract for "Data-Retention Characteristics Comparison of 2D and 3D TLC NAND Flash Memories" provides with a relation between 1time write and ~300x W/E for 2d-Nand from millenium memory towards 1yr ( https://ieeexplore.ieee.org/document/7939077 )
but, again, we don't know about controller, firmware or real-life surroundings effects with Nand only theory and laboratory conditions
Well, for component chips it's expected that this data is published, as the target audience needs to know.
But it seems these datasheet aren't publicly available? Haven't searched much, but for example, Micron seems to require one to send a "Document Access Request". Was trying to check MT29F1T08EELEEJ4-T (128GB).
Micron lost 2Bn in cash due to loss of NAND demand, and it is projected to have lower demand. I wonder how it will impact the industry. Also a shame that there's no NAND on the planet that can drive the endurance factor likes of Optane and I do not see it happening either. All the new PCIe5.0 NVMe are a joke. No improvements at all barely using the overrated PCIe bus bandwidth as the Random is far more important than the Sequential.
Plus I will mention again how the Firecuda 530 4TB NVMe is the only TLC NAND that can rival MLC technology. That beast has 5100TBW while the top end SN850X and 990 Pro max out at 1/2 of it at 2400TBW. Pathetic. I got fleeced over the 980 Pro SSD purchase last year with insane cost around $500 for 2TBx2. Today that is almost 1/2 price. And the 530 Firecuda launched at insane rip off junk at $900+ now it is 1/2 at $500 still massively overpriced. I hope the new actively cooled garbage PCIe5.0 SSDs come and reduce the price of the existing PCIe4.0 drives. I do not have hope in Samsung doing anything as their SSDs are plagued with Firmware problems which cause drive wear, lockups and other failures, both 980Pro and 990Pro. Samsung as a brand of reliability and performance has dropped off like a rock only being floated because of the brand name, if they keep pushing garbage, they are going to be irrelevant as the 3rd party Controllers with open NAND market is going to phase their in-house rubbish. Also a sin to name the TLC as Pro.
A lot of people were disappointed that the 980PRO used TLC. And yes, endurance is the last thing they think of because they want you to buy MORE so your SSD must BREAK MORE.
again, who are the 'people'? if it's enterprise, all the bean-counters and engineers care about is that the device meets its SLA/warranty lifetime. the device will be replaced at, or nearly, warranty expiration. the company gets to right off the device as an expense, so You The Taxpayer pay for most of it anyway.
only retail cares about endurance, because we're too poor to do scheduled replacement. which can be a pain in the butt.
it will be interesting to find out how those Samsung pieces of kit have faired in bidnezz.
As In, I do a lot of reading of tech articles, SSD reviews, and I talk to geeks. Next time I'll demand names and 3 proofs of identity of "everyone" so you can arrest them for "wrong think" or maybe you want me? IDK. Why not just accept that based on why I have read "people" -- not everyone -- wishes the 980 PRO would have been MLC?
The average person, in my experience, wants their devices (literally every device) to last well beyond warranty. So "bean counters" don't really factor into the viewpoint of the people I'm talking about.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
34 Comments
Back to Article
name99 - Friday, March 31, 2023 - link
Thanks, Anton. There was enough background explanation in this article to move it way beyond just a press release regurgitation.meacupla - Friday, March 31, 2023 - link
I don't have a good understanding of NAND chips.For reference, how does this one compare to other NAND chips?
400MB/s peak per chip doesn't sound that fast, because joining 8 of them together would only result in 3200MB/s. (Sequential or Random not specified)
This seems average when a Samsung 980 Pro 1TB will hit up to 7000MB/s sequential read/write.
Now, if that 400MB/s per chip was random read/write, then that would be very impressive.
FunBunny2 - Friday, March 31, 2023 - link
"Now, if that 400MB/s per chip was random read/write, then that would be very impressive."it's never been clear to me why sequential and random don't, by the arithmetic, work out to be (nearly) identical. we know, don't we?, that NAND/SSD data isn't stored contiguously as it is on disks, but spread all over the NAND, and importantly, re-spread during operation vis wear-leveling. so how does sequential access measure better?
Golgatha777 - Friday, March 31, 2023 - link
Might be the sustained data rate, and 3.2GB/sec would be amazing for that. Most of the sustained read/write fall off a cliff when hybrid SLC cache is exhausted. Some of my NVME drives still can't keep up with the sustained throughput of my 970 Pro SSD on large file transfers for instance. Probably the worst I've seen is a 1TB NVME I use for backup data that falls to about 100MB/sec sustained once the cache is exhausted.ballsystemlord - Friday, March 31, 2023 - link
That's quite characteristic of a QLC drive.saratoga4 - Friday, March 31, 2023 - link
>it's never been clear to me why sequential and random don't, by the arithmetic, work out to be (nearly) identical.Because of latency. There's a ~100us delay between when you start a read and when data is available. For 4KB reads at QD1 (meaning each waits for the one before to complete), that is 10,000 per second, or 40MB/s. Going faster means lower latency and NAND is optimized for density/bandwidth but not latency.
JTWrenn - Tuesday, April 4, 2023 - link
In short, the controller can't do as many small block pulls and set them up exactly right. It's not quite the same as seek time but it is really about the controller not the medium. If you look at high end enterprise controllers and ssd's the rates are much closer.JTWrenn - Tuesday, April 4, 2023 - link
Sorry to put a finer point on it....there is no such thing as a random read speed really. It's latency on the random then you start reading, then you do a find and read again. So random is usually not measure in read speed but rather the latency to find the new file/parts of a file and then it starts to read as fast as it can. Depending on the size of the file you run into efficiencies of file systems, operating systems, and controllers but generally it's two completely different things.erinadreno - Friday, March 31, 2023 - link
According to specification, ONFI uses 8 or 16bit data bus. This multiplies with 3.2GT and 4 channel can give you up to 12.8GBps bandwidth. Theoretically you can have a low end dramless 4 channel controller hitting that speed. Or more than pcie 5.0x4 can provide if you use an 8 channel controllermeacupla - Saturday, April 1, 2023 - link
okay, now that sounds fast.sheh - Friday, March 31, 2023 - link
Layer count comparison:https://blocksandfiles.com/2023/03/31/kioxia-and-w...
PeachNCream - Friday, March 31, 2023 - link
Great a die shrunk TLC NAND chip. I'm sure the lack of detail about P/E endurance is not at all an indicator of further reductions in useful lifespan at all. Nope, not even a little bit. Instead please pay attention to performance specs.ballsystemlord - Friday, March 31, 2023 - link
And how about retention after power-down? I've yet to hear anything about that for QLC flash nor for many of the newer TLC flash chips.sheh - Saturday, April 1, 2023 - link
@ballsystemlordPresumably, meeting the JEDEC specs. Other than that I don't recall ever seeing anyone publish detailed info on retention.
back2future - Saturday, April 1, 2023 - link
e.g.2015 technology level https://www.anandtech.com/show/9248/the-truth-abou...
2018 QLC (market entry) expectations, article endurance (TBW) and comments data retention? (year(s)) https://www.anandtech.com/show/13078/the-intel-ssd...
2021 FerriSSDs (?) https://www.eetimes.com/silicon-motions-ferrissd-h... ("At higher temperatures, NAND Flash memory cells age faster, so the duration of data retention falls faster.", but have a look at Jedec specification with increasing data retention on higher active temp(eratures) for client/enterprise (?) )
sheh - Sunday, April 2, 2023 - link
These are just models and extrapolations. Now that we've had consumer SSDs for 15 years it's time someone published concrete real life data.The Tech Report RIP did an SSD endurance test back then, but conducting retention tests is much difficult.
(If you go looking for the retention articles, brace yourself. The site was purchased a few years ago and turned into a random crypto-gambling-clickbait cesspool.)
back2future - Monday, April 3, 2023 - link
A 2020 summary for results between 3d-TLC and QLC flash ("The errors and the optimal read voltages of 3D TLC and QLC flash-memory chips are explored.") might be a starting point ( https://www.microarch.org/micro53/papers/738300a48... ) although it seems pretty theoretical results considering looking for consumer friendly review or summary on data retention parameters and storage treatment recommendations.I do agree, public data on data retention seems rare considering market share of SSDs nowadays.
ballsystemlord - Tuesday, April 4, 2023 - link
back2future & shehThanks for trying to look into this matter. I did try and research it a few years ago with about the same success you two have had.
The 2015 AT article was from one manufacturer (Intel) many many years ago and I think it was for planar flash, and before stacking, and before QLC came out, and before we changed the cell type to charge trap flash.... so not really that useful.
The 2018 QLC comment section and 2021 FerriSSDs articles are just some off hand comments and an advertisement of a product intended for the auto market. The JEDEC standard I think is JESD219A. It costs as much as a 1TB PCIe 4.0 flash drive to obtain, just so I know about how long my SSDs will retain their data -- and I probably have to sign an agreement not to tell anyone else about data-retention. When will the tech world learn to give us some hard data?! What happened to science being important?! Is it only for their exclusive club?!
The Infineon paper you cite is talking about floating gate flash which, AFAIK, is not used in modern consumer flash products. They don't test anything in their other than storing the drive at 85C, which how long it retains data is based on how many times it was P/E (Programmed and Ereased).
Regarding your IEEE article, that's the most interesting of the bunch. I'll have to read that later...
ballsystemlord - Tuesday, April 4, 2023 - link
EDIT: Regarding the Infineon paper, (AFAIK) it's 85C they're testing at, 55C real world operating conditions. I.E. cold storage isn't taken into consideration -- and as you might recall from the Intel post you cited from AT, 55C is the absolute best case scenario. Even Intel, back then, got ~8 years of data retention in that bast case scenario.See section 3.1, third paragraph from the bottom.
back2future - Wednesday, April 5, 2023 - link
2015 article is real information with look-up tables from JEDEC standard for hardware that time (if there's smth. like that for 202x QLC Nand flash comparing development of retention and hardware qualities was possible?), 2018 shows steady interest on retention, but only on commenters sections (being one example on that maybe biased view) and FerriSSDs being one example for technology dev.s to remedy short data retention specs (another was reheating flash cells having high P/E cycle numbers for curing defects and somewhat extremely extending P/E cycles, IIRC from 1-10k (~MLC or TLC flash classes) to <1-100M, but only theoretical considerations made on that approach)55°C is highest temperature for active device on JEDEC (Intel) table, enabling highest ~400 weeks data retention (on SLC/MLC types?) with steady(?)/average 25°C storage temperatures.
Me, no flash gate differences, nor production node sizes considered with researching some overview information.
Infineon and IEEC abstract correlate on information towards each other (~1k years endurance for 1time or low W/E cycles, but seems nonsense expecting linearity for a 1k years retention within diagram figure6, i'ld absolutely agree on that point) for Infineon with probably 2d SLC or MLC flash cells (10k-100k P/E endurance mentioned).
What's data retention for HDD's, current USB flash devices, sd-cards or M.2&NVME Nand storage devices (, DVD or Bluray)?
sheh - Saturday, April 1, 2023 - link
Has any other recent press release, by any manufacturer, included P/E specs?back2future - Monday, April 3, 2023 - link
only found another more theoretical approach from Infineon ~2021 "AN217979 Endurance and Data Retention Characterization of Infineon Flash Memory" ( https://www.infineon.com/dgdl/Infineon-AN217979_En... )back2future - Monday, April 3, 2023 - link
and a short view on an abstract for "Data-Retention Characteristics Comparison of 2D and 3D TLC NAND Flash Memories" provides with a relation between 1time write and ~300x W/E for 2d-Nand from millenium memory towards 1yr ( https://ieeexplore.ieee.org/document/7939077 )but, again, we don't know about controller, firmware or real-life surroundings effects with Nand only theory and laboratory conditions
sheh - Monday, April 3, 2023 - link
Well, for component chips it's expected that this data is published, as the target audience needs to know.But it seems these datasheet aren't publicly available? Haven't searched much, but for example, Micron seems to require one to send a "Document Access Request". Was trying to check MT29F1T08EELEEJ4-T (128GB).
Silver5urfer - Saturday, April 1, 2023 - link
Micron lost 2Bn in cash due to loss of NAND demand, and it is projected to have lower demand. I wonder how it will impact the industry. Also a shame that there's no NAND on the planet that can drive the endurance factor likes of Optane and I do not see it happening either. All the new PCIe5.0 NVMe are a joke. No improvements at all barely using the overrated PCIe bus bandwidth as the Random is far more important than the Sequential.Plus I will mention again how the Firecuda 530 4TB NVMe is the only TLC NAND that can rival MLC technology. That beast has 5100TBW while the top end SN850X and 990 Pro max out at 1/2 of it at 2400TBW. Pathetic. I got fleeced over the 980 Pro SSD purchase last year with insane cost around $500 for 2TBx2. Today that is almost 1/2 price. And the 530 Firecuda launched at insane rip off junk at $900+ now it is 1/2 at $500 still massively overpriced. I hope the new actively cooled garbage PCIe5.0 SSDs come and reduce the price of the existing PCIe4.0 drives. I do not have hope in Samsung doing anything as their SSDs are plagued with Firmware problems which cause drive wear, lockups and other failures, both 980Pro and 990Pro. Samsung as a brand of reliability and performance has dropped off like a rock only being floated because of the brand name, if they keep pushing garbage, they are going to be irrelevant as the 3rd party Controllers with open NAND market is going to phase their in-house rubbish. Also a sin to name the TLC as Pro.
ballsystemlord - Saturday, April 1, 2023 - link
A lot of people were disappointed that the 980PRO used TLC. And yes, endurance is the last thing they think of because they want you to buy MORE so your SSD must BREAK MORE.FunBunny2 - Monday, April 3, 2023 - link
"endurance is the last thing they think of"again, who are the 'people'? if it's enterprise, all the bean-counters and engineers care about is that the device meets its SLA/warranty lifetime. the device will be replaced at, or nearly, warranty expiration. the company gets to right off the device as an expense, so You The Taxpayer pay for most of it anyway.
only retail cares about endurance, because we're too poor to do scheduled replacement. which can be a pain in the butt.
it will be interesting to find out how those Samsung pieces of kit have faired in bidnezz.
ballsystemlord - Tuesday, April 4, 2023 - link
As In, I do a lot of reading of tech articles, SSD reviews, and I talk to geeks. Next time I'll demand names and 3 proofs of identity of "everyone" so you can arrest them for "wrong think" or maybe you want me? IDK. Why not just accept that based on why I have read "people" -- not everyone -- wishes the 980 PRO would have been MLC?ballsystemlord - Tuesday, April 4, 2023 - link
The average person, in my experience, wants their devices (literally every device) to last well beyond warranty. So "bean counters" don't really factor into the viewpoint of the people I'm talking about.coburn_c - Saturday, April 1, 2023 - link
Reads like gpt wrote itPeachNCream - Saturday, April 1, 2023 - link
If I could get away with using ChatGPT output to generate income I would certainly put it to use so I could do something else with my time.Threska - Sunday, April 2, 2023 - link
Forum management.Wereweeb - Monday, April 3, 2023 - link
It's almost like ChatGPT draws from articles written by journalists when asked to explain something...FunBunny2 - Monday, April 3, 2023 - link
"ChatGPT draws from articles written by journalists "I have it on good authority that 90% of the innterTubes is porn, 5% is Trump, 4% is militant evangelicals, and 1% the rest of us. :)