Intel High Endurance Technology NAND question

[Hulk]

I don't understand how the Intel DC3700 can achieve 10 full drive writes per day with MLC NAND. That's 18,250 writes. And if you consider the fact that this drive has about 75% over provisioning then it's still over 13,000 writes.

I thought 25nm MLC NAND was good for 3000 writes? And this is 20nm MLC NAND. Jeez, when they say high endurance they really mean it.
How is this NAND different from that in say the 520 or 330 series? Can binning amount to the endurance difference between the 520/330 and 3700? Or is it a totally different process.

BTW, I got the first question in the webcast. Yeah I guess I'm a nerd because it was exciting to hear it!

 

[Hellhammer]

The NAND in DC3700 is 25nm HET-MLC, not standard 20nm MLC. You are usually using the highest binned chips for HET-MLC but there are architectural differences as well. HET-MLC uses lower voltages for programming and erasing, which results in longer program/erase times but increases the endurance of the drive considerably. In other word, HET-MLC is not as fast as standard NAND. Data retention is also 3 months instead of 1 year, which allows the drive to be rated higher in terms of endurance.

 

[Hulk]

Thanks for the info. I didn't realize that data retention was finite on SSDs? Does this mean that if I pull my Intel 330 drive in a year I could lose the data on it?

 

[Mark R]

Thanks for the info. I didn't realize that data retention was finite on SSDs? Does this mean that if I pull my Intel 330 drive in a year I could lose the data on it?

Yes. Although it is unlikely. It also depends on the age and wear of the memory.

In NAND flash, the electrical charge stored in the memory cells gradually leaks out. However, as cells age, the leakage gets worse. When a manufacturer calculates the expected cycle life of a NAND memory, they estimate the number of cycles before the NAND no longer meets specifications (e.g. when too much data corruption occurs, or when the cells "fade" too quickly).

To an extent, the manufacturers choose their allowable level of data corruption and data retention period, and use that as the basis for their cycle life claims.for example, they aim for 1 year retention, and work out that after 5,000 cycles, the NAND will no longer reliably hold data for 1 year; so they say the max life time is 5000 cycles. When new, the same NAND might have a 10 year retention.

If you are prepared to accept a lower performance of the cells, then you can increase the number of cycles before your performance limit is reached.

So, if you simply change your spec, so that 3 months retention is allowable, instead of 1 year - then you get more cycles before the NAND fails to meet spec. This is where some of the cycle gain of high-endurance NAND comes from (but there are some design differences as well).

 

[Hulk]

Great info, thanks. I guess that means we'll be using mechanical hard disks for back-up and archiving for the foreseeable future.

 

[bononos]

Yes. Although it is unlikely. It also depends on the age and wear of the memory.

In NAND flash, the electrical charge stored in the memory cells gradually leaks out. However, as cells age, the leakage gets worse. When a manufacturer calculates the expected cycle life of a NAND memory, they estimate the number of cycles before the NAND no longer meets specifications (e.g. when too much data corruption occurs, or when the cells "fade" too quickly).

......

So do you have to keep the ssd plugged in to make sure that the memory cells are still charged up or do you actually have to rewrite the data?

 

[Hellhammer]

So do you have to keep the ssd plugged in to make sure that the memory cells are still charged up or do you actually have to rewrite the data?

As long as the SSD is powered on, the controller will "refresh" the cells if necessary.

 

[Mark R]

So do you have to keep the ssd plugged in to make sure that the memory cells are still charged up or do you actually have to rewrite the data?

Any decent SSD control processor will periodically read the data, error check and correct it, and then rewrite it.

However, it is not at all clear how quickly this happens, how long a complete "refresh" takes, or how you can tell whether the whole drive has been refreshed. This, unfortunately is a closely-guarded trade-secret of the SSD vendors. Maybe the drive only does it very slowly in the background - so that it takes 1 month to refresh the whole drive? Only the SSD design engineers know.

Flash data does "fade" very slightly during reading. So, decent SSD controllers will keep track of how much reading is going on, and which sectors are getting read the most. The controller will then expedite refreshing of heavily read areas, to ensure maximum data integrity. Possibly, by repeatedly reading the whole drive a dozen times, you might be able to trick the controller into a full refresh - but again, exactly what conditions trigger a NAND refresh are a closely-guarded secret.

Simply keeping the flash powered is generally not enough. This isn't really releveant to SSDs, as they necessarily contain a control processor; but some memory cards (e.g. XD cards) do not contain a processor, and rely on the system they are connected to to do all maintenance. Simply keeping the card powered up won't help data retention if the host system doesn't periodically refresh the data.