Over 20 times the endurance of best in class datacenter SSDs, basic rendering with very little info, m.2 form factor.
How likely is this product to ever ship?
Over 20 times the endurance of best in class datacenter SSDs, basic rendering with very little info, m.2 form factor.
How likely is this product to ever ship?
There’s a lot of ways to do this that make sense. Operating the NAND in MLC or SLC mode would be the most obvious and expensive route and should be able to easily hit those numbers, but they could also be playing creative games with the warranty period and rated operating conditions, and expecting higher replacement rates, counting on that AI MoneyNOW! to enable them to claw a market advantage elsewhere to weather the oncoming storm.
Perhaps there is also a caveat that there is no guaranteed data retention after turning the power off. As a “memory extension cache drive” only.
Could be, IIRC Optane is rated for not very long on data retention actually, which might be part of it’s high endurance rating. However, with stacked NAND, the data retention characteristics started taking a different, less linear shape, so I think this probably isn’t it.
I can’t find the papers through all the AI sludge anymore though, but unless I’m crazy, the curve favors data retention in 3D nand relatively early in it’s write life, and drops off faster after the rated PE cycle count.
That doesn’t look like the case, judging by its TBW rating of 219,000TBW.
Samsung 990 Pro 2TB has a rating of 1200TBW for reference.
But if your warranty is a year, it doesn’t matter what your TBW is, because the product is out of warranty before it ever reaches that… ![]()
Plus, you can claim it failed because oh the temperature was too high, or too low, or varied too much, or varied too little, or the space dust wasn’t aligned just right or whatever. NAND data retention and write endurance are tied heavily to temps and voltages.
According to its spec sheet:
Warranty
1. Limited 5-years or 219,000TBW.
2. Limited warranty based on 5 years or 219,000TBW, whichever comes first. (*TBW is evaluated by JEDEC workload standard. )
*TBW (Terabyte Written): Terabytes Written is the total amount of data that can be written into a SSD before it is likely to fail.
3. When the usage of an NVME SSD as indicated by the ““Percentage Used”” (SMART ID: 05) reaches 100 means out of warranty. (A new unused product will show the number of 0)
4. When the usage of an NVME SSD as indicated by the “Data units written” (SMART ID: 07) reaches “6D80000000” (Hexadecimal) means out of warranty.
Must be PMLC/PSLC then. PSLC would definitely get you there.
Or they’re just lying about it.
I think it’s a 8tb qlc ssd in SLC mode (2TB) or 4TB QLC (1TB), similar to that solidigim D7-P5810 which performs kinda close to optane, latency and DWPD wise
Wendell covered it in a video
It’ll depend on how much it costs, if the 2TB one is alittle more expensive compared to other 2TB M.2s, it’ll will be an interesting product but if it’s priced like 8TB M.2s which are crazy expensive, wouldn’t must people just get an enterprise M.2 or U.2?
Also it’s annoying it doesn’t have PLP or anything to sustain the writes, i thought AI workloads can be pretty write intensive? If the write performance will drop as soon as the DRAM cache is full, what’s the point…
If the product is legit, then I’m very interested in this product
The cache is just for holding the FTL, isn’t it?
I think PSLC will be fine for write speed.
Techpowerup seems to have more info. Also, 40% overprovisioning wow.
This thing is going to cost a lot.
Interesting… 16 KiB page size. I wonder if the fact that it’s TLC run in pSLC mode that the effective page size is actually 4 KiB after two halvings (TLC @ 16 KiB, pMLC @ 8 KiB, pSLC @ 4 KiB), matching the physical and logical sector sizes seen by the operating system. Or is this math wrong?
My evidence is that TechPowerup gives the same page size for SSDs using the same NAND chip but running in different modes: SSD Database | TechPowerUp
QLC you mean? QLC would be 16:12:8:4 for 4,3,2,1 bits per cell.
Each bit doubles the number of states vs the previous, but it’s still only storing one more bit.
It kind of looks that way to me, but I have no idea. ![]()
Where did you find your information on Optane having bad data retention?
From what I found, Optane has pretty good data retention when I was looking around at it, superior to NAND. This is because the solid state material has its physical phase changed to conduct current or not in order to be a specific bit, and it does not need to store any electrical charge to maintain this state after that (similar to hard drive changing magnetic state to store a bit). As far as I could find, on a properly working Optane drive the data retention should be many years (5-7 depending on ambient storage temp). When an Optane drive has completely expended its drive write endurance to the cells (which is far higher than NAND SSDs), only then does it become rated down to 3 months data retention when powered off.
NAND on the other hand is a floating gate that stores electricity to specify what bit it is. This gate very slowly drains electricity when not refreshed and eventually it will become corrupted data. Many people have found (myself included) that leaving a SSD unpowered for 1-2 years results in significant corrupted data, but also unfortunately if data on the drive is not written to in a cell within a couple of years it also can become corrupted even if the drive was powered on many times within those years. This is because consumer SSDs don’t auto refresh cells that have been sitting holding data automatically.
Ah right! Doubled/halved the voltage levels, but incremented/decremented bit capacity per cell.
2 KiB is still less than 4 KiB, so ![]()
Here’s to hoping that the random 4 KiB I/O doesn’t suck. I know a few places where high capacity (p)SLC would find a great home…
EDIT: unfortunate, but it seems like these are going to have 8 KiB effective page size.
Subtracting 4 KiB twice from 16 KiB yields 8 KiB.
Might be, 3DXP is really young technology. It’s really sad that we won’t see it take off in the end,
especially now that people are pitching SLC NAND as an alternative to RAM, and… That’s basically what 3DXP was made to do, and still does better…
Oops, I overwrote a post for some reason? rip
Something something data retention ratings are EOL, and not OOB ratings, Intel’s early optane data retention ratings were very low, and enterprise NAND is rated for much shorter retention than consumer NAND, implying a relation between TBW ratings and EOL data retention.
Also something about 1 being 1/3 of 3 instead of 1/2 of 3, and also I have no idea of that’s directly related to page size.
It was my understanding that Optane would retain data significantly longer than NAND despite initial conservative ratings Intel put out.
Another weird thing I’ve heard about Optane is that failure to change cell states is very common and the controller must check every cell that is written and very often has to attempt multiple re-writes of a cell before successful.
My layman’s thinking is that this highlights how stubborn the cells can be to change state.
Sadly, I just saw this post. I could have cleared this up. It’s a Phison E18 with updated LDPC technology, running in SLC mode, with the extra space rolled into overprovisioning. So an 8TB (8,192) drive in SLC is 2,730GB. The extra 730GB is OP. Pair the SLC, OP, and the modern LDPC gets you there.
Pricing has yet to be disclosed, but given the specifications, it should be cheaper than the AI TOP 100E 1TB, which retails over $2,600 at some retailers.
that price…
but the good news is they’re making a cheaper version, still with a crazy high endurance, but now im struggling to figure out why someone would buy a 2.6K$ 1TB M.2 over a U.2(which has hotswap and PLP, features that are very much welcomed in enterprise envirionments)
First note on that article…just increase the order of magnitude by one to present very high numbers. Very bad behaviour for Toms hardware. We’re at PBW as of 2023/2024 and 100PBW is nothing new.
We have enterprise SSDs with more than that. Or Optane. Or rather niche Enterprise MLC drives.
The market for that drive are people who can’t afford a server/workstation board and pay 2.6k for an M.2 drive. Or AI people in need of 100PBW drives in their laptop.
And with a DRAM cache this is more for bursty workloads (consumer drives) as write speed probably drops at some point (how to reach that 100PBW?). And no IOPS given even for the prior version. Interesting.
There is a need for really enduring and high-performing drives and storage in AI, but do these folks buy M.2?
I’d get some U.3 and call it a day. And use overprovisioning if needed.
Or just get a 1TB of DRAM/PMEM instead? infinite endurance and way faster.
Quite a bit of what you just said is very wrong.
DRAM cache SSDs don’t cache user data (your data), they cache the flash translation layer. Think of it like a map of your data. Windows and other OS have a map of your data but it is basically a map of a hard drive. SSDs don’t retain data the same way and your data doesn’t just go to a cell and it stay there. The SSDs shuffle the data from time to time to optimize performance, endurance, and data retention. Before the data map the OS has doesn’t line up with the map on the drive, there has to be a table of data keeping track of the differences. We call this the translation layer.
The AI Top SSDs are for Gigabyte’s AI Top software. You need the specialized SSDs to run LLMs that do not fit in the GPU memory. Phison developed a technology that builds a (sort of) memory pool so you can run a 70B model on lower end hardware like the Gigabyte 4070 TI Super AI Top GPUs. Before our (Phison) software, most AI programs would need to store all that data in the GPUs (roughly 1.7TB of data with a 70B model), but with the Phison aiDAPTIVLink software the data is cached on the high-endurance SSD and is swapped between the GPU and the SSD. This workload needs a high endurance SSD to be warrantied for 5 years.
As far as the performance of these drives, since they are 100% SLC (plus overprovisioning), you get consistent write performance above 6GB/s, no TLC or QLC performance drop off.