[Discuss] SSD: enterprise vs. consumer Flash

[blu at nedharvey.com (Edward Ned Harvey)]

> From: discuss-bounces+blu=nedharvey.com at blu.org [mailto:discuss-
> bounces+blu=nedharvey.com at blu.org] On Behalf Of Tom Metro
> 
> Edward Ned Harvey wrote:
> > The flash they use in USB sticks and SD cards is the same flash they use
in
> > enterrpise hard drives.
> 
> Are you sure about that?

Of course there's no way I can be sure.  I know, at work, a year or two ago
when we were designing chips for flash, the head engineer told me they were
all moving away from SLC to MLC.  I am certain there does *exist* such a
thing as higher performance or higher reliability flash versus other flash,
but I can't say where they get used.

I can certainly say that the controller matters more than the flash itself.
Even if you have cheap slow unreliable flash, as long as your controller is
doing wear leveling and error detection, you should have a fairly reliable
device.  And when you look around at cheap USB fobs, they very often boast
very high (USB3) speeds, which are totally impossible in practice, because
they're not doing pipelining or garbage collection, etc.

Once every so often, I'll buy a cheap-junk usb HDD, and a super-top of the
line USB flash drive, and compare them.  So far, the flash drive has never
even come close to the performance or reliability of the HDD.  But the same
is certainly *not* true for an SSD.  If I wanted fast external storage, I
would stick an SSD into an enclosure rather than use the fob.


> What makes an enterprise SSD controller more reliable? Or are you saying
> the controller itself isn't more reliable, but it implements hardware
> wear-leveling algorithms that makes the overall SSD more reliable?

Wear leveling, FEC, in particular.  Without it, it's impossible for you to
have a reliable flash-based storage device.  

And then, I'm sure the chips themselves may be manufactured to a higher spec
too.


> > Flash itself is dirt cheap.  What you're really buying is the on-device
> > flash controller, that maps the flash blocks to virtual HDD blocks, and
> > implements the USB/SATA operations.
> 
> Are you sure about that?

Greatly sure.  We were engaging some major producers of flash and SSD's,
attempting to sell them improved chips to do certain functions... Our folks
were able to determine (or estimate) the costs of all the components, and
the main expense was licensing the patents for the various techniques of
stuff that's done by the controller.  All the wear leveling, block
remapping, etc.


> The cost of semiconductors generally comes down to quantity of silicon
> required (die size) and factors that impact yield. Greater complexity
> increases the probability of a defect, and reduces yield. Similarly,
> cutting edge speed requirements leads to the need to test and sort
> parts, throwing out (or labeling with lower specs) the failing parts.
> This is why high-end CPUs are expensive.

In flash and SSD's, the cost of silicon is relatively small.  If you knew
how much it cost to manufacture your SSD, you would be pissed.  The money
goes to marketing & distribution, patent licensing, and simply margin.  

 
> Have you seen evidence to the contrary?

I'm intentionally vague.  Anything I could possibly know would necessarily
be under NDA.  So, no.  No evidence.