SSDs and Capacitors

[mindless1]

There have been some SSDs using a capacitor buffer to help clear the write cache during sudden power loss. It would seem to be a bad implementation, far less than optimal use of components to just add some capacitors, no matter how great their value, directly across the 5V rail.

For the purposes of this discussion I am only considering SSDs powered by only the 5V rail. My point is that directly across the 5V rail, it's back feeding to the rest of the system, and there's a limit to how much capacitance you could add before PSU-On state rise to 5V time became excessive to charge the capacitor(s), and in certain situations, the surge current too.

I am wondering what anyone out there knows about typical (or even a few examples would be a start) SSD stability/behavior during a slow voltage drop event. Would the SSD attempt to continue to function but the reduced voltage would potentially introduce errors, and/or would a sufficient amount of capacitance mean that the DRAM (or SRAM on some designs) clears the cache and then it would (or wouldn't) matter much if a low voltage unstable condition arose.

If this kind of condition happens, with some implementations of prevention it could happen every single time power is lost, like a normal system shut-off.

For this reason I propose the following and hope you can find problems or suggest an alternative that is:
1) Low cost
2) Reasonable size
3) Reliable long term, meaning over 10 years minimum.
4) Not requiring complex circuits beyond what is reasonable for DIY'ers.

I have designed fairly complex circuits and etched some, sent some to PCB houses, but it seems like a diminishing return for the time and money if the critical information about how SSDs respond to voltage dropping slowly is know.

The knee jerk reaction is supercapacitors but I've read a bit about how their ESR is too high. I thought also about powering a buck converter off the 12V rail for multiplying the effectiveness of any given capacitance.

Both of these notions lead me back towards the possibility of a simpler approach. What if you had a 5V rail and ground, and a resistor in series on the 5V rail, say 1KOhm to limit current so it isn't excessive upon PSU power-on charging a capacitor bank, then in parallel to that you had the 5V rail going to the SSD(s), with both isolated from each other by a pair of schottky diodes?

Who knows what the dropout voltage of a typical SSD is? At low current a schottky that results in only 0.2V dropout shouldn't be hard to find, but I wonder about the behavior of a typical SSD as voltage drops after the cache is cleared, especially if it's a long process, or does it not really matter at that point?

I know some of you will just state that this is a reason for a whole-system UPS. Perhaps, but this is targeted more towards a specific issue.

I'm looking more for information rather than ideas about using zener diodes, relays, transistors, and more. I've contemplated such things and come back to wondering what happens if you just throw a whole lot of capacitance at the problem so long as you have inrush current limiting?

 

[Mr Evil]

I'm not an expert on SSDs, but I expect that the ones with a backup capacitor use a dedicated controller to charge a supercapacitor slowly while the drive is operating normally. In an undervoltage condition, the controller will disconnect the supply, and start drawing power from the supercapacitor, thus it won't inadvertently try to supply power to the whole PC.

There will be a DC-DC converter in there somewhere so that the SSD never has to deal with a low voltage (as long as it finishes writing before the supercapacitor is drained).

I have seen controllers like that made by Linear Tech. They actually mention SSDs in the marketing fluff, but I don't know if any actual SSDs use those particular controllers.

However you do it, you're going to need a supercapacitor to hold enough charge to provide power for a reasonable amount of time, and if you're going to spend the money on one of them, you might as well spend a bit more and use a proper controller instead of trying to hack something together from discrete components.

 

[mindless1]

I read that a supercapacitor doesn't have low enough impedance. Intel put some capacitors in a few of theirs like a 730 series IIRC but it has 1(?)GB cache, and fairly low uF values, a pair of 47uF/35V, looked like it was more what they could fit in there than anything else but rated at a quite high voltage presumably to get the impedance down without putting a large capacitance directly on the 5V rail... just a guess, I don't expect Intel engineers to give away trade secrets but given that they could have used a higher capacitance value with a lower voltage rating in the same available space.

On the other hand a fast regulator feeding from a supercap might be able to compensate enough, except the LTC3350 you linked has the supercap on the output rather than input which makes sense to control charge current but doesn't do anything about voltage decline during discharge. Well, the datasheet for it does also state "In addition, the step-down converter can run in reverse as a step-up converter to deliver power from the supercapacitor stack to the backup supply rail.", but this seems to be within the assumption you have low voltage rated supercaps, while there should be some reasonably priced for audio amp stiffener applications but they are quite large to fit in a computer case.

I think a solution can be idealized and certainly a custom IC might do the "best" job from a performance standpoint, but it leads me back to time and expense for a one-off project, vs just throwing a lot of capacitance at it. I might be able to do testing on a specific SSD but that doesn't necessarily provide good enough data for any other model that's not nearly identical.

I'm thinking it might be necessary to involve a step down voltage regulator. Then it would draw power from the 12V rail and once it can't regulate to 5.0V (or some lower value) then it can shut down far faster than having a massive capacitor bank get all the way to 0V. It could have a diode before the regulator supply capacitor so that doesn't drain back into the system. Heck even a severely lossy linear regulator would provide that benefit but heat management seems more of a hassle than using a switcher.

Maybe I need to get some controller datasheets to see how low a voltage they can tolerate. Certainly a DRAM cache isn't directly fed 5V and flash is now under 2V isn't it? Then there's the question of whether the controller was designed to operate off 3.3V supply for 1.8" mSATA designs.

What I'm thinking is that it may not be necessary to keep supply voltage at a hard 5.0V, that a fair amount of capacitance might do the trick. I mean the cache isn't that large, it's not as though it should take even a second to clear it, right?

Someone, maybe it was Samsung, has a brochure that specs they used a 1500uF cap to get a mere 40ms or something (my numbers may be off), but being a brochure it didn't get more technical than that. Not having an o-scope I wouldn't be able to make measurements at the ms level. So, I could attempt some elaborate costly circuit and not be able to quantify the effect, or try tens of thousands uF capacitance. I am supposing that might get me a few hundred ms of viable power to the SSD, but really I had intended something for closer to a half dozen SSDs.

 

[Mr Evil]

I read that a supercapacitor doesn't have low enough impedance...

Depends. If you look at the sort of products that are available now, there are compact supercapacitors intended for backup purposes that have very high ESR (>100Ω), but there are also larger ones with low ESR (<0.1Ω).

...I mean the cache isn't that large, it's not as though it should take even a second to clear it, right?..

I went and looked at a review of the Samsung 950 Pro 512GB, which has 512MB of cache. The worst-case write performance (4KB random) would take about 2 seconds to completely empty the cache. At 3.8W of power consumption, that's nearly 8J of energy.

To store 8J, assuming a common supercapacitor voltage of 2.7V (higher voltage supercapacitors are actually multiple low voltage ones in series) and assuming that you could drain it completely, you would need a capacitance of about 2F. You could stack two 1F capacitors to get a more reasonable 5.4V, and it would cost about £0.60 ($0.80) in volume.

 

[William Gaatjes]

Depends. If you look at the sort of products that are available now, there are compact supercapacitors intended for backup purposes that have very high ESR (>100Ω), but there are also larger ones with low ESR (<0.1Ω).


I went and looked at a review of the Samsung 950 Pro 512GB, which has 512MB of cache. The worst-case write performance (4KB random) would take about 2 seconds to completely empty the cache. At 3.8W of power consumption, that's nearly 8J of energy.

To store 8J, assuming a common supercapacitor voltage of 2.7V (higher voltage supercapacitors are actually multiple low voltage ones in series) and assuming that you could drain it completely, you would need a capacitance of about 2F. You could stack two 1F capacitors to get a more reasonable 5.4V, and it would cost about £0.60 ($0.80) in volume.

I agree.
An added advantage of going to a higher voltage, is reduced current draw from the supercapacitor for the same output power with switching step down regulators. This will increase efficiency as resistive (mainly ESR) losses are lower for the same output power. Meaning that a higher percentage of the energy in the capacitors is used to provide power to the circuit in the ssd. This can result in choosing a cheaper supercapacitor while it has higher ESR and still get the required energy needed.

Also, it takes longer before the undervoltage lock out of the smps is reached.
The best smps ic would be a buck/boost converter that can do step down and step up to maintain proper output voltage.

 

[mindless1]

^ True, higher voltage lower current, BUT in this context the higher voltage would be 12V, meaning from a typical 2.7V supercap, I need an array of 2.5X as many of them, while I could just put that amount in parallel to halve the current each supplies, which seems close to a wash with respect to current though not for hold up time at whatever minimum voltage the particular SSDs need.

Further in context, I could spend hours on this which has to be worth some dollar value, or just spend that amount on more supercaps.

The dropout voltage of the SSD seems to be a critical factor. It's easy to say go with a higher voltage then buck regulate but in this scenario there is only the option of 12V instead of 5V, which with the typical supercaps being 2.7V, means less than half the capacitance per dollar. Even so it seems desirable to regulate down from 12V, except that the more involved this gets, the more I'm leaning towards just an UPS.

What I'm looking into now is SSDs without a separate DRAM cache, just a little SRAM integrated into the controller like found on models using Marvel 88NV1120. Oh, I suppose I didn't mention that I care little about performance so long as it isn't as slow as a slug.

 

[William Gaatjes]

I do not know how often you experience a power failure, but a lithium ion battery with charger and power path ideal diode from linear technology could work. A super capacitors has longer life when compared to a lithium battery.
But yes.
A super capacitor array combined with a buck boost converter with external fets (lower rdson) , it is very easy to maintain voltage long enough to do a write back.But you have to signal the ssd controller somehow that power failure is imminent and that cache to flash copy is needed.

 

[William Gaatjes]

It is waiting for F-ram inside the ssd controller or as separate chip to have a power failure safe ssd. But it would for the coming times not be megabytes. And perhaps slower.

 

[William Gaatjes]

here is news for the replacement of dram and flash.

http://www.eetimes.com/document.asp?doc_id=1330387&
Fujitsu Is Licensee of Nantero's Carbon-Nanotube RAM

Fabless chip company Fujitsu Semiconductor and foundry Mie Fujitsu Semiconductor have both announced that they are licensees of carbon-nanotube non-volatile memory technology from Nantero Inc. (Woburn, Mass.).
Nantero was founded in 2001 and has spent 15 years developing its technology, which it claims offers the potential to become a non-volatile replacement for DRAM. It also offers rewrite speeds and endurances thousands of times higher than those of NAND flash memory.
Nantero has been pursuing an intellectual property licensing business model for some time and has been claiming that its technology has been installed in numerous wafer fabs and foundries but until now has not revealed any licensees.
Greg Schmergel, Nantero's CEO and co-founder, told EE Times Europe that the Fujitsu licenses are at the 55nm node with the 40nm node set to follow. He added that Nantero is working with other partners at the 2X-nm node. Shmergel said that Nantero has more than 12 customers in place for its CNT memory technology including "several of the top ten users of memory globally."
Fujitsu Semiconductor plans to develop a custom chip with embedded NRAM by the end of 2018, with the goal of expanding the product line-up into stand-alone NRAM components. Mie Fujitsu Semiconductor, which is a pure-play foundry, plans to offer NRAM-based technology to its foundry customers.
Nantero is also supporting design work on a multigigabyte DDR4 stand-alone memory, Schmergel said. The capacity of the design and any component is not yet final but will be four to eight layers and between 8 and 32Gbits, Schmergel said. "However, we think Fujitsu could be the first to get products out in 2018," he added.
The principle of operation of the nanotube memory is a layer of carbon nanotubes (CNTs) in a random mesh matrix structure with hundreds of CNTs per device with many different intersection points, embedded within a conventional CMOS process. The number of CNTs that make contact and the effective resistance across the layer is dependent upon an applied voltage and can be set and reset.
The resulting memory offers memory cell switching speeds of the order of 20-picoseconds at low-energy together with a practical write speed of 5ns with endurance of the order of 10^11 cycles. This holds out the prospect that CNT-based NRAM can be superior to competitor technologies such as ReRAM and phase-change memory and scale better in geometry to become a universal memory to replace both DRAM and NAND flash memory.
Amongst other applications Fujitsu Semiconductor is expecting NRAM to make a convenient follow-on to ferroelectric RAM (FRAM), which struggles to scale below about 65nm.
"We will be able to build on our experience and skill in this field [FRAM] to develop and produce NRAM as well," said Masato Matsumiya, vice president of system memory at Fujitsu Semiconductor, in a statement. "The combination of Nantero's technology with our design and production capabilities promises to meet the longstanding needs of our customers for non-volatile memory that is higher density, faster, more energy efficiency, and with a higher rewrite cycle."
When asked why Fujitsu was going public on the licensing of Nantero technology Schmergel said: "I think they are ready to talk to customers, they want to engage with customers."
Schmergel added that the announcement signals the maturity and manufacturing readiness of CNT-based memory and highlights the fact that the semiconductor industry is already in the necessary transition to replace flash memory and DRAM.

 

[Red Squirrel]

That rhymes with Wun Hung Lo.

(this was in response to some spam post)

 

[Red Squirrel]

On original topic... wonder if you could get away with simply a capacitor and diode to stop it from back feeding. The rest of the computer components would be off with drive still feeding off the capacitor, thus there would no longer be any calls for new data writes, so you could technically just give the ssd enough time to finish existing writing. Of course there would be data corruption from the hard shut down, but the SSD itself would not be damaged.

The ultimate answer of course is that the whole system should be on UPS, but it seems a capacitor like this would be rather cheap to at least save the SSD should a computer with no UPS lose power or the UPS itself fails.

 

[William Gaatjes]

It would be nice to use instead of a ups, a custom build circuit that is connected between the psu connectors and the mainboard. As soon as the mains power shuts down because of some external cause, the pc gets the signal to shut down. Powered by lipo and smps and power path ideal diodes(Linear Technology) to perform a seamless switch between lipo battery and psu. I guess a small microcontroller with an ftdi usb to serial chip could be used and a small program that as soon as it gets a specific string from the usb to serial comport, signals windows or linux to shut down properly. Would be a nice project.

 

[Red Squirrel]

Hmmm kinda like a built in short run UPS. Could definitely be neat to have that as part of the standard. It could force a ACPI shutdown of the OS. Though OS makers also need to smarten up and make the shut down sequence faster. I don't know why it takes so long to shut down a typical computer.

As a custom setup, it could simply trigger the power button, which would force a ACPI shutdown.

 

[William Gaatjes]

I was thinking, For a complete pc, there is going to be some serious engineering to handle the high currents. Especially the graphics card and the cpu.
A power path ideal diode circuit with battery backup power for the ssd and hhd alone would be easier to design. But how would a hdd and a ssd know that power is about to fail ?
I assume that it could be done by dropping the voltage a bit (For example, from 3.3V to 3V and keep the voltage at 3V.) so that some undervoltage detection circuit kicks in and the ssd and hdd starts writing local cache data to the flash/platter ?

 

[agent00f]

Hmmm kinda like a built in short run UPS. Could definitely be neat to have that as part of the standard. It could force a ACPI shutdown of the OS. Though OS makers also need to smarten up and make the shut down sequence faster. I don't know why it takes so long to shut down a typical computer.

As a custom setup, it could simply trigger the power button, which would force a ACPI shutdown.

Any of the processes might have data/state (in memory, etc) which eventually needs to be written to disk, which can take up to any amount of time. Some kind of mini-ups on the SSD per this thread solves a small part of the problem, there is no general catch-all solution.

 

[gbeirn]

On original topic... wonder if you could get away with simply a capacitor and diode to stop it from back feeding. The rest of the computer components would be off with drive still feeding off the capacitor, thus there would no longer be any calls for new data writes, so you could technically just give the ssd enough time to finish existing writing. Of course there would be data corruption from the hard shut down, but the SSD itself would not be damaged.

The ultimate answer of course is that the whole system should be on UPS, but it seems a capacitor like this would be rather cheap to at least save the SSD should a computer with no UPS lose power or the UPS itself fails.

That was my original thought, one could easily make an inline SATA adapter that houses a few capacitors, of course a diode to prevent current from flowing back to the system when power is cut.

The controlller continues to write the data in its buffer and with nothing new coming in, power could easily be lost without corrupting that specific write.

 

[NeoPTLD]

Transfer switch.
http://www.intel.com/content/dam/ww.../ssd-power-loss-imminent-technology-brief.pdf

Page 2.

this explains how you don't lose the power back into the rail.