At last month's FAST (File and Storage Technologies) conference, researchers from UCSD and Microsoft Research presented a paper titled "The Bleak Future of NAND Flash Memory" that's put the rosy all-flash-all-the-time future predicted by some in doubt for the long term. The gist of the paper's argument is that the physics of flash, and the researchers' testing, indicate that as flash density increases, the life and performance of flash will degrade to the point that by 2024 flash will no longer be a viable solution.
The researchers used a PCIe flash card and 45 flash chips from different vendors with different cell sizes to extrapolate the performance of SLC, MLC (2 bit per cell) and TLC (3 bit per cell) flash as cell geometries shrink from the current 32 to 34nm and 24nm cell geometries to the 6.5nm cells that could be on the market in 2014. They discovered that as cell geometries shrank, latency doubled and the bit error rate more than tripled for MLC. Of course increased latency means lower IOPS, and for a fixed-size SSD they predict an almost 50% fall off in performance.
Things are worse for TLC, where latency reaches a very disk-like 2.5ms and better for SLC. I'm frankly skeptical that TLC flash will ever make it even in the laptop SSD market. Storing eight different charge levels, and maintaining guard bands among them, will result in too high a bit error rate as the flash controller can't tell the difference between a 4 and a 5 in a cell, as well as endurance well below a thousand erase cycles.
Some of the reports I've seen have made the mistake of conflating SSDs with flash. While the vast majority of SSDs today are based on flash memory, with a small number of DRAM-based SSDs making up the rest, several technologies are waiting in the wings to replace flash when it runs out of steam. Just as flash has replaced everything from UV-erasable EPROM to bubble memory phase change memory, ReRAM based on HP's Memristor technology and/or electron spin transfer torque memory should reach the market in the next 10 years, just as the flash roadmap reaches its bleak future.
Even if entirely new technologies aren't ready in time, the report didn't investigate the benefits of sophisticated flash controllers like those from SandForce, now owned by LSI, and Anobit, now owned by Apple, that store additional ECC information on the flash. Smarter controllers should give us a few more years.
DeepStorage has no commercial relationship with USCD (though I do occasionally have UCSD Pascal flashbacks) or Microsoft, other than paying for software. DeepStorage is, however, still investing in flash SSDs.
