Pure Storage says SSDs will struggle to get past 30 TB capacities because of DRAM limitations.
Shawn Rosemarin, Pure VP for R&D, told Blocks and Files that, generally speaking, a commercial off-the-shelf (COTS) SSD needs 1 GB of DRAM for every 1 TB of raw NAND capacity. “So 30 terabyte drive, 30 gigabytes of DRAM, 75 terabyte drives, 75 gigabytes of DRAM, so on and so forth. You can imagine when you start to look at something like a 75 terabyte drive with 75 gigs of DRAM, that’s roughly as much DRAM inside an individual drive as there is in servers today.”
He reckons there are three side effects of this. “We know that DRAM fails significantly more than NAND. We know that the price, or the cost, I should say, of DRAM is significantly higher than the cost of NAND. We know that the energy efficiency, the actual draw of DRAM, is significantly higher than that of NAND.”
DRAM is needed for the Flash Translation Layer (FTL) software that runs as firmware in the SSD’s controller. This allows you to write incoming data to different physical flash pages no matter which logical block the data was intended for. The DRAM holds the FTL mappings and metadata for this and things such as the space capacity set aside for over-provisioning.
Generally speaking, the FTL has to do more work as the number of bits per cell increases to cope with shorter cell write cycle numbers by minimizing write amplification. As SSD capacities increase, all of this is an added cost that will make up a larger and larger portion of the overall SSD media cost.
In effect, as SSDs use more DRAM, Rosemarin reckons their cost will rise, they will need more electricity, and they will become less reliable.
The cost of DRAM has not declined at the same rate over the past few years as NAND. This may affect coming SSDs. Rosemarin told us: “When you look at these SSDs … that require all this DRAM, do you have a significant issue around the cost of being able to facilitate bringing all that DRAM into those devices?
“The only way around this is to ultimately find a better crutch, or a better solution [than] DRAM, one that is more cost effective, less energy consuming.” We need to “essentially move that capability that today is being delivered via DRAM directly into software, which is what we’ve done at Pure with Direct Flash Modules. We don’t have that reliance on DRAM” at the drive level.
Any such DRAM cost increase will not constrain Pure’s Direct Flash Module (DFM) capacity roadmap. The DFMs are effectively just a collection of flash chips with FTL done at the system-wide level in Pure’s controller and its software, instead of at the SSD level as in other all-flash array storage systems.
He thinks this means “when you take the DFM, our delivery mechanism of capacity, we think it provides a much, much faster innovation cycle over the next few years” than that of COTS SSDs.
In other words, Pure’s DFMs will increase capacity much faster than off-the-shelf SSDs.
“So when you think about what we’ve done with our 75 terabytes, and what we’ve told the market, we’ll do with 150s and 300s. And by the way, we have every intention to scale beyond 300, to 600, and even to 1.2 petabyte per DFM.”
Pure is talking about 150 TB DFMs coming in 2025 and 300 TB ones possibly in 2026.
Bootnote
IBM is another storage supplier with proprietary flash drives, FlashCore Modules.
