SSD write endurance/durability

[It's Not Lupus]

Should I be worried about TLC and smaller nm NAND regarding write endurance/durability? I suppose over-provisioning with help with this. I have bought the Crucial M4 and Samsung 830 128GB before the next gen gets out as the NAND just gets smaller and smaller. And these drives are fast enough for me. How impressive would it be to still have drives working after a decade or two.

 

[Cerb]

For now, I'd just avoid TLC. Leave it to uses that might have short lifespans, like thumb drives.

After that, it's an open question. Historically, the write cycles have gone down. Intel and Toshiba are claiming to be holding the line, at least, for now, but what the future holds, we really don't know.

 

[Coup27]

It all depends on what you do with the computer. I would be quite happy to put a TLC SSD in my Mum's computer because all she does is browse, office and general stuff.

I would also probably be happy putting TLC in all of works computers as all of the heavy data lifting is handled by the server.

TLC would not be suitable for a machine that handles a lot of data.

People have been and will continue to slag off TLC because it isn't as good as SLC or MLC. That's not what it's trying to do. It is offering a cheaper cost option where the usage pattern would would suit it.

The whole subject of NAND endurance has been blown out of all proportional by people who simply do not understand it.

 

[bryanW1995]

It all depends on what you do with the computer. I would be quite happy to put a TLC SSD in my Mum's computer because all she does is browse, office and general stuff.

I would also probably be happy putting TLC in all of works computers as all of the heavy data lifting is handled by the server.

TLC would not be suitable for a machine that handles a lot of data.

People have been and will continue to slag off TLC because it isn't as good as SLC or MLC. That's not what it's trying to do. It is offering a cheaper cost option where the usage pattern would would suit it.

The whole subject of NAND endurance has been blown out of all proportional by people who simply do not understand it.

I think that a better way to say that would be, "Intel/etc are kicking the crap out of NAND currently, rendering longevity issues moot for almost all users". There is no guarantee that this trend will continue, though I admit that I'm betting on the engineers' ability to improve longevity over NAND's ability to be intractable/short-lived.

 

[Cerb]

People have been and will continue to slag off TLC because it isn't as good as SLC or MLC. That's not what it's trying to do. It is offering a cheaper cost option where the usage pattern would would suit it.

But how much cheaper is it? So far, not nearly enough. SLC is expensive, and its benefits miniscule for 99% of users. The reduced cost of MLC is substantial, and likely to increase over time.

Unless TLC can provide complete SSDs for half the cost of MLC, or less, I don't see where it is sufficiently advantageous. With current controllers and WA, TLC could bring us into having to worry about lifetime writes, which we mostly don't have to do on PCs, with MLC.

I have no way to guarantee that anyone's usage pattern fits, without recording it, making, "the usage pattern fits," bit quite difficult to match up with confidence. Current WA on MLC amounts to, "if you don't run a high-disk-utilization server, your WA should be much less than 2." By the time anyone not running a high-disk-utilization server eats up their writes on a current-gen drive, the drive will be obsolete and small. With TLC, those 1.5-2 WA uncommon cases could be quite damning for the drive's longevity, even if the common case is closer to 0.7 (not uncommon with light document-heavy workloads and SF drives).

 

[Hellhammer]

The whole subject of NAND endurance has been blown out of all proportional by people who simply do not understand it.

Especially when people only look at P/E cycles and write amplification. If the claims of adaptive DSP are even remotely realistic, then the raw P/E cycle count doesn't mean anything. For example STEC is claiming that they can achieve over 60k P/E cycles with regular 3-5k MLC NAND.

 

[VirtualLarry]

For example STEC is claiming that they can achieve over 60k P/E cycles with regular 3-5k MLC NAND.

That sounds promising, but how is that possible? I thought that the interfaces to NAND didn't allow "raw" access to the sense lines, or whatever they use, I thought it just accepted binary data, and the NAND chip itself performed the program/erase. No?

 

[Hellhammer]

That sounds promising, but how is that possible? I thought that the interfaces to NAND didn't allow "raw" access to the sense lines, or whatever they use, I thought it just accepted binary data, and the NAND chip itself performed the program/erase. No?

NAND is controlled by a voltage sent by the SSD controller. When a drive is manufactured, certain voltage parameters are used. However, as the NAND is used, these parameters may change (the silicon oxide wears out which leads to cell leakage, trapping etc). The idea of adaptive DSP is to monitor NAND characteristics and modify the voltage parameters based on the changes. Without DSP, you may be applying multiple voltages before getting the desired action; and each time you apply a voltage, the silicon oxide wears out more.

I think DSP definitely has its use but I'm rather skeptical about the presented figures.

 

[Cerb]

I think DSP definitely has its use but I'm rather skeptical about the presented figures.

I'll bet they're highly unrealistic, using best-case scenarios. But, if they can even get a 2-3x minimum improvement in p/e cycles, that could get us 5+ years, assuming slightly reduced 'normal' p/e cycles for smaller NAND coming up.

More importantly, I think, for the likes of Intel/Hitachi, Micron, Toshiba, and Samsung, it could mean using top bins of MLC only for their high margin drives, which will be much cheaper than SLC. Not that we really care, for the most part, but it would be worth investing in such R&D for long-term server drive profits.