Intel gets out of flash SSD business

[porina]

Quotes

Quote

In a brief news release from Intel this afternoon, the chip firm has announced that it has closed on the first stage of its deal to sell its SSD business to SK hynix. As of today, SK hynix has now formally acquired the bulk of Intel’s NAND and SSD businesses, as well as the company’s NAND fab in Dalian, China. Intel will continue to hold a small stake until 2025, and in the meantime Intel’s former SSD assets have been spun-off into a new SK hynix subsidiary, Solidigm.

Sources

https://www.anandtech.com/show/17134/intel-sells-ssd-business-to-sk-hynix-as-new-subsidiary-solidigm

 

Summary

Intel gets out of the flash SSD business, selling it to SK Hynix who will run it as a new subsidiary Solidigm. Optane is not included in this deal, and is retained by Intel.

 

My thoughts

While they had some success with the 600p, which at the time provided great capacity/cost ratio, it does feel to me this is an area where to be most effective you need to go large or go home. Selling this off allows them to focus on other areas of business. Don't think it'll make much if any impact to the consumer/enthusiast space which remain served by other established brands.

 

Optane remains a question as to what Intel does with it. The last time I looked they didn't have a way to manufacture it any more, as they parted ways from Micron who later sold off the fab that was making it.

[HQuan]

Optane not included? Phew almost skipped a heartbeat of mine.

 

[dizmo]

This is very, very old news.

 

https://www.google.com/amp/s/www.theverge.com/platform/amp/2020/10/20/21525701/intel-nand-ssd-business-sk-hynix-9-billion-optane-memory

[RejZoR]

1 hour ago, dizmo said:

This is very, very old news.

 

https://www.google.com/amp/s/www.theverge.com/platform/amp/2020/10/20/21525701/intel-nand-ssd-business-sk-hynix-9-billion-optane-memory

They were selling it back then. Now it's official as sale was finalized now. Hence the news.

[Middcore]

4 hours ago, HQuan said:

Optane not included? Phew almost skipped a heartbeat of mine.

 

 

With Micron bailing on the partnership earlier this year, I wouldn't be optimistic about the future of Optane. 

[HQuan]

1 hour ago, Middcore said:

With Micron bailing on the partnership earlier this year, I wouldn't be optimistic about the future of Optane. 

What? Micron bailing with Intel earlier this year? I never hear this new before! (because I was on a vacation I guess...)

But Micron still a good brand!

[Middcore]

Just now, HQuan said:

What? Micron bailing with Intel earlier this year? I never hear this new before! (because I was on a vacation I guess...)

But Micron still a good brand!

https://www.architecting.it/blog/micron-3dxpoint-intel-optane/

 

https://www.extremetech.com/computing/320932-micron-ends-3d-xpoint-development-will-sell-its-optane-fab

 

https://blocksandfiles.com/2021/03/17/intel-optane-memory-future/

 

https://searchstorage.techtarget.com/news/252499716/Intel-breaks-silence-on-effects-of-Microns-3D-XPoint-exit

[porina]

7 hours ago, RejZoR said:

They were selling it back then. Now it's official as sale was finalized now. Hence the news.

Yup. As a parallel, you could compare the "old" news as like nvidia stating an intention to buying Arm. If they ever get the deal through, which is looking ever less likely, that'll be new news again.

 

So what we have on this thread is the first part of the execution of the transfer of assets to SK Hynix. There's still another phase later on, but that's more a formality I guess and less newsworthy.

 

5 hours ago, Middcore said:

With Micron bailing on the partnership earlier this year, I wouldn't be optimistic about the future of Optane. 

Without looking it up, I thought it was even earlier than this year. Intel and Micron agreed not to continue their relationship on Optane a while back. This year I think it was Micron moving it up a notch by actively trying to sell the fab that was making it. As such, we don't know that Intel has any other source for it. They have the rights to make it themselves but I'm not aware of any statement they have done so.

[tim0901]

8 hours ago, Middcore said:

With Micron bailing on the partnership earlier this year, I wouldn't be optimistic about the future of Optane. 

 

2 hours ago, porina said:

Yup. As a parallel, you could compare the "old" news as like nvidia stating an intention to buying Arm. If they ever get the deal through, which is looking ever less likely, that'll be new news again.

 

So what we have on this thread is the first part of the execution of the transfer of assets to SK Hynix. There's still another phase later on, but that's more a formality I guess and less newsworthy.

 

Without looking it up, I thought it was even earlier than this year. Intel and Micron agreed not to continue their relationship on Optane a while back. This year I think it was Micron moving it up a notch by actively trying to sell the fab that was making it. As such, we don't know that Intel has any other source for it. They have the rights to make it themselves but I'm not aware of any statement they have done so.

Intel make Optane themselves in their New Mexico fabs. This likely wasn't enough capacity to keep the consumer stuff around though - hence they dropped it to focus on the server space where it makes more sense as a product.

 

From what I can tell though it isn't exactly a profitable product (hence Micron bailed ship) so who knows how much longer it will stick around.

[porina]

9 hours ago, tim0901 said:

Intel make Optane themselves in their New Mexico fabs.

Thanks, I hadn't seen that stated previously. So far I've only managed to find one Tom's Hardware article stating such. I did see many other references saying R&D on Optane was being done there.

[leadeater]

3 hours ago, porina said:

Thanks, I hadn't seen that stated previously. So far I've only managed to find one Tom's Hardware article stating such. I did see many other references saying R&D on Optane was being done there.

Intel Fabs labelled as R&D can be a little deceptive as they can be used to production purposes as well, where it makes sense.

[whm1974]

Maybe Intel could more with Optane as a really fast SSD with no wearing out. You know instead of as Memory Modules. Think PCIe X4 to X16 SSD Cards. Or M.2 for Notebooks.

[leadeater]

3 hours ago, whm1974 said:

Maybe Intel could more with Optane as a really fast SSD with no wearing out. You know instead of as Memory Modules. Think PCIe X4 to X16 SSD Cards. Or M.2 for Notebooks.

They did, cost too much and nobody wanted it due to that.

[whm1974]

1 hour ago, leadeater said:

They did, cost too much and nobody wanted it due to that.

Yeah but it seems that neither Intel nor Micron made very much of it to begin with. I can a PCIe Card from x4 to x16, or even x32 using Flash will wear out if using solely for Swap.

[leadeater]

51 minutes ago, whm1974 said:

Yeah but it seems that neither Intel nor Micron made very much of it to begin with. I can a PCIe Card from x4 to x16, or even x32 using Flash will wear out if using solely for Swap.

They didn't make much because they knew it wasn't going to be high volume sale. Wouldn't matter if they made more the price was still way too high for what anyone was willing to pay. Also good flash you won't wear out and you shouldn't be relying on swap either, no matter how fast the storage swap is slow simply due to how it works and is utilized.

 

Like there is no point spending even more than a 980 Pro when you'll never wear out the 980 Pro.

[whm1974]

6 minutes ago, leadeater said:

They didn't make much because they knew it wasn't going to be high volume sale. Wouldn't matter if they made more the price was still way too high for what anyone was willing to pay. Also good flash you won't wear out and you shouldn't be relying on swap either, no matter how fast the storage swap is slow simply due to how it works and is utilized.

 

Like there is no point spending even more than a 980 Pro when you'll never wear out the 980 Pro.

Mass Production lowering price over time. Granted it would still cost  a bit more then Flash but not as much like now.

[leadeater]

6 minutes ago, whm1974 said:

Mass Production lowering price over time. Granted it would still cost  a bit more then Flash but not as much like now.

It was never going to get close enough, the limited capacity was to make it appear less bad however if you go by $/GB there was no chance.

 

$2.4/GB vs $0.23/GB, simply impossible to get close no matter how much you make of it.

[whm1974]

4 minutes ago, leadeater said:

It was never going to get close enough, the limited capacity was to make it appear less bad however if you go by $/GB there was no chance.

 

$2.4/GB vs $0.23/GB, simply impossible to get close no matter how much you make of it.

How expensive was Flash Memory when it was invented? From reading it was only used where you couldn't use FDD or HDD due to damaging environmental conditions. Aerospace, Military, and Spacecraft. Even that took awhile to replace Bubble Memory.   

[leadeater]

9 minutes ago, whm1974 said:

How expensive was Flash Memory when it was invented? From reading it was only used where you couldn't use FDD or HDD due to damaging environmental conditions. Aerospace, Military, and Spacecraft. Even that took awhile to replace Bubble Memory.   

Probably just as costly however that's not as relevant as you think it is. Early flash was expensive because the fabrication for silicon based storage medium was entirely new where as Optane is just a new technology based on the same underlying fabrication technology that is already mature and not going to get significantly cheaper.

 

If Optane gets cheaper than Flash gets cheaper so it's a forever losing battle Optane vs Flash $/GB. In the same way SLC can never be cheaper than MLC and MLC can never be cheaper than QLC.

[porina]

I got a 280GB 900p in November 2017 for £400 (£1.43/GB). Now for something to compare with in a similar time frame? I also got a 960 Evo (upper-mid range NVMe) in July the same year. That was £205 for 500GB (£0.41/GB). That puts the Optane about 3.5x the cost of flash in that specific example. In my head, I have a typical figure of Optane being 4x the cost relative to decent flash at the time. If I were to get the equivalent Pro at the time, that was perhaps 50% more expensive than the Evo, so obviously would close the gap between flash and Optane further. If you care about performance, which you do if you're at all seriously looking at Optane, then high end flash is the distant 2nd choice to compare with.

 

The problem for Optane consumer affordability was that over time it didn't really drop much in price, where flash did. So if you look at later flash, it got more affordable relatively speaking. So Optane remained a niche for those who genuinely needed the performance at any cost, or the few curious but deep pocketed people like myself who could get one to play with regardless. With my latest system, I just went with a 980 Pro and Optane isn't really an option now.

 

7 hours ago, leadeater said:

Also good flash you won't wear out and you shouldn't be relying on swap either, no matter how fast the storage swap is slow simply due to how it works and is utilized.

 

Like there is no point spending even more than a 980 Pro when you'll never wear out the 980 Pro.

There are some niches where it can make sense. Some day I'd still like to try for some compute world records, but it is somewhat pay2win. They're almost entirely ram capacity limited for any mortal. As a starting point I'd like a 8TB ram system (single logical box) which could run the task out of ram and not have to hit slow storage at all (except for backups/checkpoints). As far as I know that doesn't really exist, and even if it did, I can't afford it. Using and optimising swap is what people in this space actually do. Then the eternal debate: SSD vs HD. Tiered storage beyond ram and bulk doesn't work as it can be seen as a hard enough workload that caches will be quickly overwhelmed and useless. HDs have predictable performance and don't wear out from writes, although you'd need so many of them for bandwidth, statistically some will fail and you have management considerations from that. Flash SSDs will wear out unless you overprovision massively, as the workload approximates to 50% reads and writes which can go on for weeks or months. Other problem is dynamic cache SSDs would be poor performance. Again unless you overprovision massively you would be working in their cacheless performance region almost all the time. Still far cheaper than ram though. Optane would be perfect in this use case as "cheaper than ram" while not dying as quickly as flash.

 

Maybe Optane is still competitive in cost per endurance? Going back to the earlier example:

900p 280GB: 5.11 PBW (effective 18250 cycle life)

960 Evo 500GB: 200 TBW (effective 400 cycle life)

 

Ball park 25x endurance for the 3.5x cost.

[leadeater]

32 minutes ago, porina said:

Flash SSDs will wear out unless you overprovision massively

But that's the thing, they really don't. Buy appropriately and flash wear just is not a problem within the practically useful lifespan of the SSD. We have 5 year old SSDs in our Netapp that have been used for MSSQL databases, all iSCSI storage to multiple physical server clusters. None of the SSDs are below 99% health with zero reallocated or bad sectors, or anything.

 

The majority of people vastly over estimate their write loads, the rest buy low end SSDs and complain when they fail. Anyone else not covered just buys an SSD and doesn't have any problems at all unless they are in the ~1% fairly standard electronics failure rates.

 

32 minutes ago, porina said:

There are some niches where it can make sense. Some day I'd still like to try for some compute world records, but it is somewhat pay2win. They're almost entirely ram capacity limited for any mortal. As a starting point I'd like a 8TB ram system (single logical box) which could run the task out of ram and not have to hit slow storage at all (except for backups/checkpoints). As far as I know that doesn't really exist, and even if it did, I can't afford it. Using and optimising swap is what people in this space actually do.

Well my comment was specially buying Optane Storage over Flash because of fear of write endurance which statistically there isn't anything that actually backs this up in any meaningful way. Theoretically yes Optane does have higher write wear but never make the assumption this will be a problem in the first place and for the love of everything don't not buy an Intel 600p SSD because it has low wear and buy an Optane SSD instead when there is a wealth of cheaper and larger capacity Flash/NAND SSD that would suit the workload just fine and have less limitation due to the small capacities Optane comes in.

 

Also your scenario is why Optane DIMM exists not Optane Persistent Storage.

 

Optane persistent storage make sense in a very few situations where you want the performance rather than the endurance, due to the higher performance demand the endurance becomes useful as more writes are happening for example.

 

32 minutes ago, porina said:

Maybe Optane is still competitive in cost per endurance? Going back to the earlier example:

900p 280GB: 5.11 PBW (effective 18250 cycle life)

960 Evo 500GB: 200 TBW (effective 400 cycle life)

 

Ball park 25x endurance for the 3.5x cost.

Putting aside buying a storage device for swap being a really bad idea itself you'll have a much more useful component if you simply buy a 960 Pro instead, no reason to buy a 960 EVO here. Or better yet get a Ent/DC SSD or simply buy a larger 960 Pro since a 1TB provisioned at 280GB will more than likely have the same write endurance as the 900p anyway.

 

What I'm saying is don't buy Optane just because it has higher endurance if you have no idea if you actually need it at all, swap is not automatically going to mean you require it and additionally make sure there is actually a performance difference using Optane as swap over NAND in the actual applications because odds are there isn't.

 

I cannot over state just how much NAND wear is an overstated hype myth supported by dubious and inappropriate purchasing decisions or poor quality or under rated NAND SSDs. If people could stop trying to pull 40 ton semi-trailers with off the lot standard 4x4 utes/trucks then they would probably have a better towing experience 

[porina]

 

29 minutes ago, leadeater said:

But that's the thing, they really don't.

Please keep that in context of what I wrote:

Quote

 Flash SSDs will wear out unless you overprovision massively, as the workload approximates to 50% reads and writes which can go on for weeks or months.

It is a very specific use case, not a general statement your shorter quote made it look.

 

29 minutes ago, leadeater said:

What I'm saying is don't buy Optane just because it has higher endurance if you have no idea if you actually need it at all

For normal use like most around here, I fully agree that flash endurance is nothing to worry about. The only SSDs I have that have had any noticeable wear are low capacity (120GB). They're mostly just used for an OS and little more and will probably survive the life of the rest of the system. If data rot doesn't kick in first, but that's another story.

Edited January 2, 2022 by porina
Improved readability.

[leadeater]

23 minutes ago, porina said:

Please keep that in context of what I wrote:

I did, even though the origin context was a very broad "but Flash wears out" which is a statement fraught with so many problems and the reason I addressed it.

 

23 minutes ago, porina said:

It is a very specific use case, not a general statement your shorter quote made it look.

Having 50% write load, percentage wise, still does not mean that NAND cannot service the workload. Unless you specifically know what the write amount is in total TB per day and know there are no NAND SSDs on the market or due to the requirement of capacity to meet the necessary endurance that would make Optane the more economic choice.

 

Time and time again I hear endurance thrown out, zero times anyone been able to mathematically prove it would have been a problem because they don't even know how much data they would be writing per day in the first place. I hate arguing imagined problems because there is no way to resolve it, there can never be a resolution because there isn't any data to work off of.

 

Optane is 10 DWPD, NAND comes in a range of options as low as 0.1 and as high as greater than 10 DWPD so unless you are getting 10 DWPD with Optane cheaper than NAND then the endurance argument is simply a non-starter. Additionally as you said yourself you can simply buy a lower DWPD NAND based SSD of a larger capacity and over provision it to equivalent or potentially more.

 

Also I don't know why anyone wouldn't break up a computation job to consume less than 8TB if they can, but assuming for whatever reason it cannot as those cases do exist this isn't a measure of how much data was written to a device over a period of a day. Active memory and memory access as well as swap utilization isn't such a simple thing if you want to pull out a data point like data actually written and where it was written to.

 

Hard data speaks the loudest, it's really that simple.

 

I buy stuff just to see how it works out or if I'm interested in it, even with zero justification for it and that is fine, nothing wrong with that. I just don't like unjustified or unqualified reasoning as to why another product is bad to support that purchase. There doesn't have to be a solid slam dunk reason to buy something over another or a serious shortcoming of all other options on the market.

Edited January 2, 2022 by leadeater

[porina]

5 minutes ago, leadeater said:

Having 50% write load, percentage wise, still does not mean that NAND cannot service the workload.

The use case was SSD as ram, which will be a significant reduction in performance compared to ram. Implication being it is sustained at SSD best transfer rate.

 

5 minutes ago, leadeater said:

Time and time again I hear endurance thrown out, zero time anyone been able to mathematically prove it would have been a problem because they don't even know how much data they would be writing per day in the first place. I hate arguing imagined problems because they is no way to resolve it, there can never been a resolution because there isn't any data to work off of.

http://www.numberworld.org/y-cruncher/guides/swapmode.html

 

Look down the bottom of that page for "A note about SSDs". It is out of date, but similar principles still apply. I wouldn't go for the Pi record myself but considerations are very similar for others.

 

5 minutes ago, leadeater said:

Also I don't know why anyone wouldn't break up a computation job to consume less than 8TB if they can

I'm not smart enough to say how it works, but suffice to say smarter people than me haven't found a way to do it. See it like threaded software. If it were that easy to scale, you think we would have done so in more cases by now.

 

8TB is only an example amount I picked for my chosen test as something that isn't insanely expensive. For harder records it would be far more than that.

[leadeater]

32 minutes ago, porina said:

The use case was SSD as ram, which will be a significant reduction in performance compared to ram. Implication being it is sustained at SSD best transfer rate.

 

http://www.numberworld.org/y-cruncher/guides/swapmode.html

This doesn't sound like it's utilizing OS swap? Is that the case?

 

Quote

Internally, y-cruncher uses raw disk I/O. The data in this scratch buffer goes directly to/from disk via DMA. This has the following performance characteristics:

• The per-drive buffer size is the largest block of data that can be processed at a time.
• Processing each block of data requires up to 3 disk seeks. (One for the block, and one for each of the boundaries due to read-modify-write.)

I won't get in to the whole why they are using the term swap but that is not swap from what I am reading.

 

/swap or pagefile.sys and it's utilization is an entirely different thing.

 

Quote

The operating system page file swapping policies are optimized for "normal" applications. They perform terribly for specialized programs like y-cruncher. Unfortunately, it is difficult to quantify how much slower because nearly all attempts to run a ram only computation using more memory than is physically available leads to so much thrashing of the pagefile that the system becomes unresponsive. (I call it the, "Thrash of Death".) Typically the only way to recover is to hard shutdown the machine.

 

Why can't this be done with pagefile hints? Because it would be more complicated than just doing the disk I/Os manually. Furthermore, there would be no guarantee that it would actually behave as desired. In short, the OS pagefile is a black box that's better not to mess with.

 

To summarize, y-cruncher wants nothing to do with the OS pagefile. It wants the OS to get hell out of the way and let y-cruncher run unimpeded. Windows is pretty good at allowing this with the right API calls. But not in Linux. Starting from v0.7.1, y-cruncher will try to lock pages in memory to prevent the OS from entering the Thrash of Death.

Exactly as I said, hence why from what I am reading they do not use it.

 

Buy Optane in place of memory, like was mentioned, is simply not a good idea as swap does not lead to good application performance no matter how fast the storage the swap lives on.

 

Quote

With these assumptions, let's calculate how long the 500 GB SSD will last:

Writes Until Failure = (500 GB) * (5000 P/E cycles) / (2x write amplification) = 1,250,000 GB = 1.25 PB

 

Time to Failure = (1.25 PB) / (500 MB/s) / (1/2 portion of time doing writes) = 5,000,000 seconds = 1,389 hours = 58 days

Unfortunately this is not correct so I wouldn't use this at all. This is literally the type of FUD that I mean when talking about having to dispel all these NAND wear fears.

 

And the problem is still there, how much data is being written each day to storage. nowhere is that actually stated. Does y-cruncher actually fully sustain maximum disk utilization 24/7?

 

Quote

Assuming that the computation is I/O bound, it will be doing disk I/O around 80 - 95% of the time at full bandwidth. To be on the safe-side, let's assume it to be 100%.

This is a terrible assumption as an array of SSDs may very well have more I/O performance than computation could demand of it and thus not fully sustain every single SSD all the time. There's simply a serious lack of data, evidence and reasoning.

 

Also y-cruncher seems like one of the worst places to use Optane if you want to do anything serious they are talking about, you know how much 100TB of Optane would cost? I think I'll pass on that one, even for NAND heh.

 

Quote

In fact, Samsung said that they have a 128GB 850 Pro in their internal testing with over eight petabytes (that is 8,000TB) of writes and the drive still keeps going, so I tip my hat to the person who is able to wear out an 850 Pro in a client environment during my lifetime.

https://www.anandtech.com/show/8216/samsung-ssd-850-pro-128gb-256gb-1tb-review-enter-the-3d-era

 

Of course that is unofficial and not their manufacturer backed warranty.