750GB per package.

A single SSD may have anywhere from 1 to $alot of packages on the board, hence 10TB SSDs.

Note that you can query ark.intel.com to find every chip that supports ECC:

http://ark.intel.com/search/advanced?ECCMemory=true&MarketSegment=DT

A newer study just came out saying it was only 77.4%.

Erases can fail, but that's typically a gross failure in the peripheral circuitry and not a cell-level/array-level problem. It's no different than you being unable to erase your data if you have a mechanical failure on a rotation drive.

Your most likely "leakage" case is with a grown defect or a change in the flash translation layer, however, the specs are written so those old locations must be erased by a secure erase command. I know that based on NAND physics, if you do that erase, the data is gone and never coming back. IMO, there really aren't enough electrons in a charge well to reliably encode "additional" information about the prior state of a bit following an erase.

An NSA hack is always possible where they install rogue firmware on the drive that doesn't actually secure erase properly, but that kind of argument/speculation is outside the scope of my answer.

Dual-E5 Xeon systems will get you 80 lanes of Gen3 PCIe.

NVMe was invented to work around controller bandwidth and latency issues that you mention, thus getting you full "link" speed into your database

That was true over 3Gbit/s SATA perhaps, but hasn't been true for a while now.

A single PCIe SSD in a 2.5" form factor can move sequential data at 2.5GB/s or better (Gen3 x4). Modern rotating drives cap out somewhere around 200MB/s if I remember correctly, even on 6Gbit/s SATA or SAS links. That means that for sequential IO, you need 12 rotating drives to match the performance of one SSD. For random reads (700K/s on the latest samsung) you need 5800 rotating drives (assuming 120 IOPS from the rotating disk). Note that would require a queue depth of almost 700,000 which is impossible from an application standpoint. In reality, your pool of 5800 rotating drives will be MUCH slower.

At the largest of the large sizes, hard drives will likely stay behind rotation for another decade when only considering cost.

However, if you don't need terabytes of fast storage, we've already crossed the threshold where SSDs are cheaper.

Smallest hard drives you can buy these days are $60 new and store 500GB. That same $60 gets you a 128GB SSD from a "Tier 1" manufacturer (Samsung, Intel, Micron/Crucial, Sandisk, Toshiba)

For just about any storage application that fits in ~100GB or less, SSDs are both cheaper and more reliable TODAY than rotating drives.

That 100GB crossover threshold with the cheapest rotating drives will double every year or so, since today's rotating drive prices are almost completely based on the cost of the electronics and a single head/single platter mechanical system. You can't make a rotation drive significantly cheaper than today, but with each generation of SSD they can halve the number of NAND packages, shrink the PCBA, build controllers with fewer channels, etc.

If they're following industry standards, the typical guarantee on client drives is that your data will be available for 1 year of 60C bake assuming you've already cycled it to the endurance limit.

If your temperature is cooler or you haven't used up all your cycles yet, then retention will be longer than the guarantee.

Enterprise drives trade retention for endurance, thus allowing them to support more write cycles within their warranty periods. The trade-off is that the endurance limit at 60C becomes 3 months instead of 12 months, which for typical enterprise applications is more than enough. Note that when powered up but idle, the drive is performing management of the NAND, so the retention numbers are all assuming that the drive is powered off and in storage.

I'd mod you up if I had points.

Intel's X25-M was introduced in late 2008 at $1000 for 80GB ($12.50/GB), and was hard to get demand was so high.

You can currently buy enterprise-grade SSDs from multiple vendors for about $0.65/GB, with failure rates that are a fraction of 1%, and they're ten times faster (random IO) than the X25-M was.

The Numonyx venture was specifically for NOR flash manufacturing.

IMFT (Intel Micron Flash Technologies) is the NAND partnership between Intel and Micron.

3rd party barrel/chamber? Hard to blame glock for that one.

I have another problem with it.

Anyone who cares about their own safety doesn't carry a firearm without a chambered round. If you think that in a real incident you'll have time to draw and rack the slide to chamber a round, you're mistaken.

Yea. The obesity epidemic in the US correlates strongly to the publication of our food pyramid recommendations by the USDA.

Prior to that, obesity affected a very small number of people in the US.

I'm guessing it was ineffective because most people don't realize how much sugar has been added to the foods they buy, and even relatively small amounts of added sugar break the premise of "slow carb" diets.

That doesn't matter. The power supply surrounding the socket/pads will account for whatever Vcc needs to be.