Why doesnt RAM wear out like an SSD?

[Smoblikat]

If RAM and SSD are both solid state then why doesnt RAM wear out like an SSD does? by wear out i mean run out of writes. Is it because SSD is NAND and ram isnt?

 

[zanemoseley]

Good question, I'm interested in the response as well.

 

[Ben90]

DRAM for this threads purpose acts more similarly to logic transistors. The xtor gets switched back and forth depending on the value.

Flash relies on a floating gate. Basically electrons tunnel through an insulator to store the value for a decent amount of time. This tunneling behavior is what ultimately destroys the flash.

 

[razel]

Yes to your last statement. Being able to maintain data without power comes at a cost.

 

[exdeath]

DRAM uses normal capacitors that are directly connected to the circuit and are constantly discharged and recharged. Solid state semiconductor capacitors don't really have any appreciable life cycle limit. Of course being ordinary capacitors, and very tiny ones, they drain and lose their contents in nanoseconds when unpowered.

Flash uses a floating gate transistor that permanently stores a charge in a totally electrically insulated floating gate that isn't connected to anything. Hot carrier injection or tunneling is used to force electrons though the insulator and into the floating gate at very high voltages where they are essentially trapped forever (until reprogrammed again) and electrostatically cause the non floating gate to be on or off depending on the content of the charge trapped in the floating gate. Thus representing one or zero.

The high voltage and tunneling of electrons through the insulator breaks down the dielectric insulator with each erase/program cycle as electrons tunnel through the material and change the molecular state and electron composition. Basically the high voltage tunneling poisons the insulator with free electrons (?) until the dielectric starts becoming a conductor and progressively leaking so much that the floating gate can no longer hold a isolated charge. It just conducts and drains instantly through the insulator when you try to fill it and can't ever be a "1" again.

The thinner the insulator (eg smaller die feature size = higher density) the less time it takes for the insulator to be destroyed, hence the declining write cycles with each process shrink.

In higher capacity and second and third generation SSDs it's no longer a concern due to the industry now understanding things like wear leveling, write amplication, etc. A device with 120 GB and 10,000 rewrite cycles would have to write 120 GB of data before each block had a write count of one, rince and repeat 9,999 more times, and you end up with decades of typical use.

 

[greenhawk]

DRAM for this threads purpose acts more similarly to logic transistors. The xtor gets switched back and forth depending on the value.

From memory, that is close. DRAM (Dynamic RAM) is, at it's core, a small capacitor. This actually discharges very quickly (even when not used) but the memory controller is designed for this so it is continually reading and re-writing information to refresh it (though new days I think just accessing a data bank on the RAM is enough, been a while).

This design is cheap to make and has high density (so lots of memory per chip).

Another kind of ram used is SRAM, or Static RAM which is xor based, but does loose information when power is lost. While expencive to make (needs more silicon to have a single bit) it is very fast and does not suffer the need of being refreshed. It was what was commonly used for level 2 and level 3 cache of a cpu. Now it is just inbuild to the CPU core.

The SSD uses a sort of flash memory IIRC but have not bothered keeping too close attention to it.

edit: too slow it appears

 

[exdeath]

DRAM cells are 1T1C and very high density.

SRAM cells are 6T arranged in a flip flop, thus lower density, and power hungry even in CMOS, as you have a lot of transistors switching when the state changes. Lots of current draw due to the way CMOS creates a direct short from source to ground as complimentary pairs of N and P channel transistors are simultaneously on as they both switch at the same time. But oh so very fast.

I dream that 1T1M STT-MRAM will make some progress. You get the non volatile of Flash, the speed of SRAM, the feature size and density of DRAM/Flash, and near infinite rewrites and unpowered data retention of over 30 years. Can we say 256 GB of MAIN RAM that never loses contents, and eliminating the hard drive completely?

Instead of RAM + HDD you just have a single memory that is both RAM and HDD at the same time. It will be a complete paradigm shift on the way we think of RAM and HDD space. Fill it up and don't have enough free for use as RAM? Well you don't need "free RAM" to load anything into, you just access the original copy at it's permanent storage location in the non volatile storage. The entire file system including running processes and all are in "RAM" at all times even when powered down. Want to load something? No need, it's already in RAM, just execute a JMP instruction. No loading, no caching, no anything, everything is just always there instantly addressable by the CPU at multi gigabyte per second SRAM speeds... eg: transfer a folder with 2 million 4K files in about .5 seconds? We would finally be CPU limited again for EVERYTHING for the first time since CPUs reached 10 MHz. Even with a 8 core 4 GHz CPU.

I feel that with the IO bottlenecks that plague every aspect of our technology since the invention of the first magnetic tape/disk recorders, and the ever increasing size of data in our data driven world today, that the arrival of this sort of "universal memory" will be a revolution in computing that would dwarf the invention of the transistor or microprocessor itself.

I personally would see about a 2000% productivity increase if I never had to stare at another HDD access light, hourglass, progress bar, or "please wait" banner ever again.

 

[corkyg]

In simple terms, RAM is volatile and SSD is non-volatile. Reasons stated above.

 

[Mark R]

It's due to the construction. Fundamentally, DRAM
Consists of a capacitor connected to a charging circuit and a voltage sensor (transistor). The capacitor can be charged to a specific voltage, and will retain that charge for a short period. (DRAM needs a refresh circuit which reads the ram and then rewrites the data, to deal with the capacitor charge fading. Normally this is done about 4000 times per sec). Without recharging, the capacitor charge will fade.

Capacitors do not "wear out" to any significant extent, so this circuit should operate for a prolonged period.

Flash memory deals with the capacitor discharge problem by making no connections to the actual capacitor. Without any electrical connections, the capacitor won't leak, so doesn't need refreshing

But if there are no electrical connections, how is the voltage sensed? The answer is by placing the capacitor directly next to a special sense transistor that directly senses the electric field from the capacitor - this combined unit is called a floating gate transistor.

So how is the capacitor charged if there are no direct connections? This is done by using a high voltage to force electrons through the insulation around the capacitor. Under the influence of high voltage, electrons will 'punch through' (the technical name is 'tunnel' through) the insulation, to flow in or out of the capacitor. The problem is that the tunnelling process damages and degrades the insulation around the capacitor. After a few thousand cycles, so much damage accumulates, that the capacitor gets leaky and the data starts fading.

 

[notposting]

exdeath, I like your vision of the future.

 

[greenhawk]

Can we say 256 GB of MAIN RAM that never loses contents, and eliminating the hard drive completely?

we can dream.

I remember back about the 486/pentium days some company out of Germany (IIRC) made a motherboard that had 64MB of SRAM for main memory. NFI the cost, but at the time, I think 8MB of DRAM was "standard" / all you need.

I remember thinking "That would be fast!!"

 

[Anarchist420]

Slc drives were said to be better than mlcs but don't ask me why.

 

[taltamir]

It just conducts and drains instantly through the insulator when you try to fill it and can't ever be a "1" again.

Awesome writeup. Just one thing of note...

From what I understand.
Drained = 1
Full of electrons = 0

So it can never a 0 again and will always show a 1.

 

[exdeath]

It varies widely I’m sure and depends on your perspective. I know with NOR Flash that I've worked with recently, from the software programmer side as far as bits are concerned, erasing a cell by reprogramming it resets all the bits to 1, and writing to a cell drains it and turns it to a 0. You can’t set a 1 again; you have to erase the whole block to set all bits to 1. When the cell wears out you can't erase it anymore, it just stays 0. Whether this means the floating gate is + or - is unknown from the programmer's perspective.

It could go either way as far as the charge carriers go; electrons on the floating gate electrostatically repel the control gate causing the transistor to be on or off depending on N channel or P channel and then again depending on active high or active low logic convention. All a matter of vantage point I suppose.

 

[exdeath]

Slc drives were said to be better than mlcs but don't ask me why.

MLC is not fully on or off all or nothing; the FGMOS is treated as a linear device which can store various levels of charge in between. The detection circuitry has to be very sensitive to detect various levels of charge. In order to store two bits per cell, 4 discrete levels of charge which are very very small and very close to each other in magnitude must be detected precisely.

Insulator degradation is a gradual thing that occurs over time and starts to affect the cells long before the cell would be considered worn out and unuseable. If I had to guess, I'd say that due to the higher sensitivity requirement of MLC that small amounts of leakage that would not affect SLC would create false data values and data corruption with MLC much sooner. Thus MLC is rated for less rewrites than SLC. Where SLC might start at 100% and leak down to 75% and still be a 1, MLC would read what should be ‘11’ as ‘10’ (where 0-25% = ‘00’, 26-50% = ‘01’, 51-75% = ‘10’, 76-100% = ‘11’ for example.

It’s not that the physical insulator wears out faster than SLC, it’s that the threshold for how much wear affects memory reliability is much lower for MLC due to the increased sensitivity requirement. Thus MLC is rated with lower rewrite counts.

Total guess.

 

[oslama]

so a ssd will wear out sooner than hdd ?

 

[greenhawk]

Slc drives were said to be better than mlcs but don't ask me why.

SLC vs MLC?

it is "Single Level Cell" and "Multiple Level Cell". It refers to the amount of information stored in a single cell of memory. The SLC is faster as it has only a single set point to determine a 1 or a 0. Because of this, it has a lower data density per area of silicon. MLC is cheaper as each cell contains more data. This is done by using better detectors. In a SLC you only need to work with a high charge or no charge. MLC needs to be more accurate and work with more levels, so no charge, some charge, more charge, fully charged. This extra complexity leads to more data per silicon area (cheaper) but needs more circuits to support it (so slower).

SLC gives speed with price, MLC gives value and slower speed. In theory MLC is proberly going to suffer from ware faster as it is more critical that the injected "value" may change (has a smaller distance to travel between voltage levels and so implied data/information).

Given the cost difference though, not much the average consumer can do as IIRC SLC was nearly double the price per size of a MLC unit. Now with mass production kicking in, MLC is getting cheaper faster than SLC.

 

[greenhawk]

so a ssd will wear out sooner than hdd ?

hard to say as too many variables.

A hdd could be find, then the motor gives out. The media of a HDD can be exposed so the coating of the platters will react and loose it's ability to retain the magneticly stored information. The HDD could be bumped and the read/write head could contact the platter (heads use the air generated by the rotating disk to "float" a very small distance from the actual surface).

Over time both can loose the information even if nothing is done with the drive (ie: electrons leak from the cells in a SSD, magnetic field in a drive interacts with other magnetic fields (ie: other tracks) degrades to the point the read/write head can not see what it is. This degrading magnetic fields is a problem with tape storage as the tape is stored wrapped around itself.

at some point, assuming no physical or electrical damage to a SSD or HDD, the information will need to be read and written again to restore the method of recording. Sure, this is not common, but it can become a issue if using the drive as a offline backup for a decade or so. SSD's I am not so sure about as I do not know anyone rich enough to use SSD's as a long term backup method. In the normal daily use of a computer, the design of the device should not see this sort of issue for the usable life of the computer.

 

[martixy]

Well exdeath's post was illuminating. It also reminded me that there's actually a reason why they tried to teach me semiconductor technology in university.
Given the close relation of electricity and magnetism(you know, with them being basically one and the same fundamental interaction), I'm surprised some similar effect hasn't been discovered sooner. Recently we've really started to delve into quantum mechanics for the effects that drive our technology. I like where this is going.
I expect great things from the next decade.

 

[taltamir]

so a ssd will wear out sooner than hdd ?

actually, no. It will last much much longer. Its just that we can actually can measure the lifespan of an SSD vs a HDD where we have to guess based on previous experience (which shows it to be much less then the estimated lifespan of an SSD). People are just freaking out because they can see the lifespan ticking down on the meter...

The exception being some VERY specific usage scenarios (high traffic server)

 

[exdeath]

so a ssd will wear out sooner than hdd ?

Nope. You are more likely to have firmware glitches, shoddy controllers, etc locking out access, or just a bad flash chip, just like brand new HDDs can drop like flies. At least SSDs can claim teething pains from being new tech; HDDs are 5 decades old and have no excuse.

A dead SSD has nothing to do with flash wear. This should be obvious from the many people posting that their SSD just died but SMART data was 100% clean the day before with no bad sectors.

You might wear out a SATA 3 device in a "few months" if you wrote 600 MB sec every single second sustained for a "few months", assuming it doesn't go into thermal shutdown in an hour. To put into perspective how absurd that is for a standard user, if you figure max internet speeds of < 10 MB sec, you could torrent to your hearts content 24/7 for "few months" x 60 = decades of use. If torrents aren't going to kill your SSD, temp files and cookies sure aren't.

And if you do wear it out? So what, you have a backup, and you buy a $100 1 TB SATA 5 SSD to replace it.