But back when we had shortage in HDD production a few years ago, it was because many companies used factory based near some catastrophe. Here it's "just" silicon, are there only so few fabs all over the world that are able to manufacture these?
Basically, yes. Fabs are hugely expensive and they can only pump out so many wafers. The article is saying that it looks like demand is outstripping the capacity of the fabs to increase supply.
Last one - so if a fab can produce wafers at say 14nm - is there something else that can be stopping it from producing 14nm NAND flash? I mean apart from the quality.
I imagined fabs like 3d printers. I thought the toughest part is the design, and once you have it, there's now plenty of places where you can print it.
> is there something else that can be stopping it from producing 14nm NAND flash? I mean apart from the quality.
Well, you'd need the machines for that – there’s only a handful of companies able to build the EUV lithographic machines for that, most of them small German companies with a few dozen to hundred employees.
Or rather, it takes years. Of course they can increase production eventually...
But the silicon factories cost billions each, so no-one is going to scale up because of a temporary shortage. The same, even more so, for the factory-factories.
The fabs don't want to increase production of current 'old' technologies and are transitioning to smaller die processes and ramping up 3D production so there's a perfect storm of increasing demand and fab refurbishments reducing availability.
In logic, smaller is faster: less capacitance, shorter channels, and thinner oxides means faster switching.
In memory, smaller means less space to store your bits. I remember a quote from somewhere that with TLC at 20nm, you could lose only 1 ELECTRON PER YEAR from the floating gate to meet the memory retention spec.
Also, NAND had smaller feature sizes than logic for some time, before they really hit the wall. They innovated a lot of advanced techniques like quad patterning for memory cells. It helps in memory that you are optimizing for one and only one cell.
Flash fabs are different than logic fabs (that's about the extent of my understanding so I won't try to explain why). Also, logic fabs may not have much idle capacity either.
Design for stuff like nand includes a bunch of process steps that have a huge impact on the devices. NAND might require different equipment from what is already being used. It all requires intense calibration and many engineers with quite a lot of experience.
Intel planned (don't know if they ended up doing it) to convert a plant built in 2010 to memory fabrication and it was going to cost 3.5 billion dollars and take years.
High density chip fabrication processes are not commodiy services. Each foundry's process has its own design rules that your chip needs to take into account. And that's for logic. 3D NAND fabrication requires different materials and methods than logic fabrication, and each of the four manufacturers has their own design with its own manufacturing challenges. At least one or two of them have had to make significant adjustments in their fundamental memory cell design between their first announced prototypes and the shipping products that have layer counts high enough to offer competitive cost/GB.
Feast/famine cycle--it goes all the way back to DRAM in semiconductors (it's why Intel got out of DRAM).
Demand for memory spikes (new application, got cheap enough, etc.)
All the companies begin constructing new fab lines simultaneously (because, if you are more than a few months late when a new technology comes online, your competitors get all the nice, fat initial profit margins while you get the nasty bulk profit margins).
Fab lines take a while to build, but demand is still increasing. Prices spike.
Fab lines come online. Supply skyrockets. Prices crash.
Now that capacity has increased dramatically and prices are low, people start finding new uses for the cheap memory.
Demand slowly builds. Prices recover slowly.
Suddenly, somebody finds a new application for memory. Demand spikes.
Out of serious curiosity, what are these new apps that are using more and more RAM, at least for the last few cycles?
It seems like 8 GB has been the default amount for most consumer PCs for years now. This is in contrast to 15-20 years ago where PC manufacturers would really double ram every 2 years.
Step #1, observe generation migration from disk to flash, especially in the under $100 where minimum disk prices give a huge advantage to flash
Step #2 invest in production facilities that take 2 years to build
Step #3 watch the price/bit curve carefully to optimize your return on investment
Step #4 Have note7 ship by a million (with 32-128GB each)
Step #5 recall them all and throw them away
Step #6 recall the replacements
Step #7 have apple (finally) double the size (16-32GB) of the minimum size iphone 7
Step #8 have 1TB SSDs finally hit price points that enthusiasts and professionals would consider
So basically those investing 2 years ago... invested wisely. After all it's not like demand driving prices higher is bad for those who produce flash.
The article claims consumer USB storage should be the first to suffer the price increases, but a quick plot of amazon prices shows the opposite, if anything: prices steadily decreasing.
I guess I should check back in January and see if the prices rose?
Maybe this is an artificial shortage. Because it is getting cheaper to manufacture these products now. Maybe something new coming out & we need to pay for it, at least the R&D.
The article claims that SMR drives -- Shingled Magnetic Recording -- is slower than normal HDDs.
The truth is different. Unless you're unlucky enough to have a drive without a flash write buffer, SMR drives manage similar read rates to normal disks while having vastly higher write rates. For a while, that is; once the write buffer is full, it'll drop through the floor.
SMR is bad for RAID, and it's bad for very high throughput writes, but a well-made disk may actually be faster than the traditional kind for normal users while having higher capacity. See Samsung's archive disks for a nice example.
You may now ask yourself what's the point, when flash is even faster. Anyone with enough data that it won't fit on flash should probably use RAID of some kind.
And there's something to that... But they won't, and you know it, so they might as well have the capacity.
3D NAND is struggling, but it's not a complete failure. Intel/Micron 3D TLC NAND is available in volume at a great price but it's slower than Samsung's 3D TLC. Micron still hasn't shipped any 3D MLC. Samsung's 48-layer 3D NAND performs well but supply is still pretty tight and some product lines are close to a year overdue for upgrading from their 32-layer 3D NAND. Toshiba/SanDisk(Western Digital) 3D NAND is being used in iPhones but is not yet sold in SSDs; it's not clear if this is due to raw supply limitations or if they're having trouble with performance or endurance. SK Hynix has been shipping an insignificant amount of 3D MLC with sub-par density, but no 3D TLC yet.
Everybody has promises for a future of scaling up the layer count of 3D NAND to deliver big improvements in price per GB. But they're not offering much to reassure us that those newer designs will be free from the issues that have been holding back this year's planned advances.
There were never enough NAND capacity for SSD to replace HDD. And not likely until 2020 or later.
And with all the NAND manufacture consolidation you will expect NAND prices to raise all the way to 2018. Samsung, Toshiba, Sandisk/WD, Hynix combine have more then 90% market shares. i.e They pretty much control the market price.
Things will only get better once the Chinese have their Fabs up in 2018. They have poured 10s of billions in. These capital intensive business suit them well when they have no idea where to put their money.
33 comments
[ 90.9 ms ] story [ 1225 ms ] threadLowercase "as" seems compliant with the current AP recommendations: "Capitalize the principal words, including prepositions and conjunctions of four or more letters." http://writingexplained.org/ap-style/ap-style-composition-ti...
Could be a feature!
I imagined fabs like 3d printers. I thought the toughest part is the design, and once you have it, there's now plenty of places where you can print it.
Well, you'd need the machines for that – there’s only a handful of companies able to build the EUV lithographic machines for that, most of them small German companies with a few dozen to hundred employees.
And they can’t just increase production.
But the silicon factories cost billions each, so no-one is going to scale up because of a temporary shortage. The same, even more so, for the factory-factories.
In memory, smaller means less space to store your bits. I remember a quote from somewhere that with TLC at 20nm, you could lose only 1 ELECTRON PER YEAR from the floating gate to meet the memory retention spec.
Also, NAND had smaller feature sizes than logic for some time, before they really hit the wall. They innovated a lot of advanced techniques like quad patterning for memory cells. It helps in memory that you are optimizing for one and only one cell.
Intel planned (don't know if they ended up doing it) to convert a plant built in 2010 to memory fabrication and it was going to cost 3.5 billion dollars and take years.
It's not plug and play.
Demand for memory spikes (new application, got cheap enough, etc.)
All the companies begin constructing new fab lines simultaneously (because, if you are more than a few months late when a new technology comes online, your competitors get all the nice, fat initial profit margins while you get the nasty bulk profit margins).
Fab lines take a while to build, but demand is still increasing. Prices spike.
Fab lines come online. Supply skyrockets. Prices crash.
Now that capacity has increased dramatically and prices are low, people start finding new uses for the cheap memory.
Demand slowly builds. Prices recover slowly.
Suddenly, somebody finds a new application for memory. Demand spikes.
Lather, rinse, repeat.
It seems like 8 GB has been the default amount for most consumer PCs for years now. This is in contrast to 15-20 years ago where PC manufacturers would really double ram every 2 years.
For the time being, Moore's Law still seems to be operating for flash. But I doubt it's got many more iterations left.
So basically those investing 2 years ago... invested wisely. After all it's not like demand driving prices higher is bad for those who produce flash.
http://camelcamelcamel.com/Corsair-Flash-Voyager-Slider-256G...
http://camelcamelcamel.com/Patriot-512GB-Supersonic-Magnum-P...
The article claims consumer USB storage should be the first to suffer the price increases, but a quick plot of amazon prices shows the opposite, if anything: prices steadily decreasing.
I guess I should check back in January and see if the prices rose?
>We began reporting on the first signs of the looming NAND shortage all the way back in May
For your first link, May is when the flashdrive was at its lowest point followed by the price jumping up from ~$70 to ~$100.
The truth is different. Unless you're unlucky enough to have a drive without a flash write buffer, SMR drives manage similar read rates to normal disks while having vastly higher write rates. For a while, that is; once the write buffer is full, it'll drop through the floor.
SMR is bad for RAID, and it's bad for very high throughput writes, but a well-made disk may actually be faster than the traditional kind for normal users while having higher capacity. See Samsung's archive disks for a nice example.
You may now ask yourself what's the point, when flash is even faster. Anyone with enough data that it won't fit on flash should probably use RAID of some kind.
And there's something to that... But they won't, and you know it, so they might as well have the capacity.
If 3D storage isn't working, why this HN article today?[1]
[1] https://news.ycombinator.com/item?id=13107550
Everybody has promises for a future of scaling up the layer count of 3D NAND to deliver big improvements in price per GB. But they're not offering much to reassure us that those newer designs will be free from the issues that have been holding back this year's planned advances.
And with all the NAND manufacture consolidation you will expect NAND prices to raise all the way to 2018. Samsung, Toshiba, Sandisk/WD, Hynix combine have more then 90% market shares. i.e They pretty much control the market price.
Things will only get better once the Chinese have their Fabs up in 2018. They have poured 10s of billions in. These capital intensive business suit them well when they have no idea where to put their money.